AMD Outpaces Nvidia in Linux Gaming Tests as RTX 50 Drivers Falter

Gamers Nexus

Gamers Nexus testing reveals that AMD GPUs often beat Nvidia under Linux due to severe driver and frame-pacing bugs, though rapid operating system drift prevents reliable long-term benchmarking.

AI-assisted summary of RIP Windows: Linux GPU Gaming Benchmarks on Bazzite — original source. Check the original for full context.

Source published: 2025-11-25

Key ideas

Gamers Nexus conducted a multi-week benchmarking evaluation of modern GPUs on Linux using the immutable gaming-focused distribution Bazzite, assessing both raw performance and the feasibility of building an ongoing Linux GPU test suite. (0:00)
AMD Radeon GPUs exhibited significantly higher frametime consistency than Nvidia RTX 50-series cards across several titles on Linux, even outperforming the flagship RTX 5090 in games like Starfield at 1080p and 1440p due to driver regressions. (28:38)
Direct comparisons between Windows and Linux benchmark numbers are technically invalid due to architectural differences in capture hooks (MangoHud vs. PresentMon), patch versions, and translation layer execution. (5:27)
Rapid upstream software drift across kernels, Proton updates, drivers, and daily shader cache micro-updates makes long-term incremental GPU chart appending currently non-viable on Linux without periodic full suite re-tests. (45:45)

Detailed notes

## Background Published by Gamers Nexus on YouTube on November 25, 2025, this video documents the outlet's initial exploration into GPU gaming benchmarks on Linux using the Bazzite operating system. Following thousands of viewer requests, the piece focuses on establishing rigorous testing methodology, evaluating operating system viability as an alternative to Windows, and detailing the unique technical friction encountered during multi-GPU lab testing.

## TL;DR Gamers Nexus presents an extensive evaluation of gaming performance across Nvidia, AMD, and Intel GPUs running on the immutable Linux distribution Bazzite. The benchmarks reveal that while AMD hardware frequently enjoys tighter frame pacing and unexpected competitive parity against Nvidia flagships in titles like Starfield and Black Myth: Wukong, Nvidia GPUs often encounter severe frametime excursions, driver instability, and software bottlenecks under Linux translation layers. Intel's Arc Battlemage cards deliver respectable mid-range stability and ray tracing performance, though they suffer from outright game-launch failures in select titles. Ultimately, the team concludes that while Linux gaming has matured into a viable Windows alternative for general users, the rapid software drift, lack of automated tooling, and frequent regressions make long-term, incremental GPU benchmarking exceptionally difficult.

## Atmosphere & style The presentation carries Steve Burke's trademark clinical, exhaustively methodical, and dryly skeptical delivery. It feels less like an enthusiast celebration of an open-source alternative and more like an engineering autopsy of software friction, defined by rigorous transparency around anomalies, crashed test runs, and pacing spikes. The tone blends cautious optimism about escaping Microsoft's telemetry with the palpable exhaustion of lab technicians wrestling with shader compilation delays, broken native ports, and shifting software stacks.

## Bulletin - [00:00](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=0s) Gamers Nexus conducted a multi-week benchmarking evaluation of modern GPUs on Linux using the immutable gaming-focused distribution Bazzite, assessing both raw performance and the feasibility of building an ongoing Linux GPU test suite. - [28:38](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1718s) AMD Radeon GPUs exhibited significantly higher frametime consistency than Nvidia RTX 50-series cards across several titles on Linux, even outperforming the flagship RTX 5090 in games like Starfield at 1080p and 1440p due to driver regressions. - [05:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=327s) Direct comparisons between Windows and Linux benchmark numbers are technically invalid due to architectural differences in capture hooks (MangoHud vs. PresentMon), patch versions, and translation layer execution. - [45:45](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2745s) Rapid upstream software drift across kernels, Proton updates, drivers, and daily shader cache micro-updates makes long-term incremental GPU chart appending currently non-viable on Linux without periodic full suite re-tests.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=0s) Microsoft's aggressive telemetry, data harvesting, and forced online connectivity are accelerating community demand for viable Linux gaming alternatives. - [07:29](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=449s) Gamers Nexus selected Bazzite as the standardized test operating system due to its immutable architecture, responsive maintainers, and gaming-focused out-of-the-box configurations. - [08:36](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=516s) Bazzite maintainers removed gaming-oriented kernel optimizations in August after discovering they caused 2% to 30% performance regressions on Intel hybrid-core CPUs. - [05:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=327s) Testing on Linux requires an entirely manual workflow because automated 10-year Windows test harnesses cannot run natively on the platform. - [05:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=327s) Direct cross-platform chart comparisons between Windows and Linux are methodologically flawed due to differences in metric capture hooks between MangoHud and PresentMon. - [21:49](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1309s) In Black Myth: Wukong, Nvidia RTX 50-series GPUs suffered severe frametime pacing spikes from 8 ms to 20 ms, while AMD and Intel GPUs maintained steady 3 to 4 ms intervals. - [28:38](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1718s) Starfield on Linux exposed major Nvidia driver flaws, allowing AMD's RX 9070 XT and RX 9070 to beat the RTX 5090 at 1080p and 1440p resolutions. - [34:45](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2085s) The native Linux Vulkan build of Baldur's Gate 3 produced erratic, unrepeatable results on Nvidia hardware, while the Windows DirectX 11 build under Proton provided consistent frame pacing. - [37:24](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2244s) In ray-traced workloads such as Cyberpunk 2077, Nvidia maintained its raw performance advantage over AMD, while 8 GB graphics cards routinely crashed due to VRAM exhaustion under Linux. - [28:38](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1718s) Intel Arc Battlemage cards exhibited tight frametime pacing across functional titles, but failed to launch Starfield entirely and crashed with ray tracing enabled in Cyberpunk 2077. - [39:28](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2368s) Kernel-level anti-cheat remains a critical barrier for Linux gaming adoption, forcing the removal of titles like F1 24 from the test suite. - [44:05](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2645s) MangoHud benchmark logging introduced significant instability during testing, including frequent game crashes directly upon concluding a capture run. - [45:45](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2745s) Maintaining a rolling Linux GPU benchmark dataset is currently unfeasible because asynchronous updates across the OS, Proton, and shader caches introduce untracked variance over time.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=0s) Benchmarking Gaming on Linux and the Push Away from Windows

Microsoft’s increasingly aggressive telemetry and data harvesting have dramatically lowered the barrier of resistance for users looking to leave Windows behind. Between the growing enshittification of the operating system and the explosive popularity of SteamOS, audience requests for Linux gaming benchmarks have expanded far beyond a handful of dedicated niche enthusiasts. Steve Burke emphasizes that while Microsoft has built an impressive business model around surveillance and data acquisition, it has created a massive incentive for the PC gaming community to seek an escape hatch. To answer those calls and establish baseline data for players considering a platform switch, Gamers Nexus embarked on an extensive, multi-week research and benchmarking effort.

This first round of GPU testing represents an exploratory trial by fire. The team spent two to three weeks upfront researching bench hygiene, testing controls, and environment isolation, followed by three weeks of non-stop benchmarking on the Bazzite operating system. Steve notes that while he is not a first-time Linux user, his last experience daily-driving the OS occurred between roughly 2009 and 2012, primarily dealing with network-attached storage since then. Throughout the testing phase, the team consulted directly with the maintainers of Bazzite to properly dial in their testing variables, ensure clean data sets, and understand the best practices required to generate trustworthy comparisons.

To finance this labor-intensive new testing initiative, Gamers Nexus directly funded the project through their store merchandise rather than vendor sponsorships. Highlights include their in-house designed retro-themed drink coaster packs—featuring four distinct Blender-modeled designs such as an old-school PC front panel with a floppy disk drive, a classic game controller, an IBM Model M-inspired numeric keypad, and a GN dial-and-ridge motif—alongside an "Unhinged" vaporwave smiley T-shirt tied to their upcoming investigative journalism into a UK raid concerning Sega and Nintendo development hardware.

> “It’s a really impressive business model they've built, which is spying.” — Steve Burke

- The benchmarking project required roughly 5 to 6 weeks total: 2 to 3 weeks of preliminary isolation research and bench hygiene, followed by 3 weeks of continuous testing. - The project is fully crowdfunded via viewer support, including retro hardware-themed drink coasters and investigative special report fundraisers.

---

[03:25](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=205s) Realistic Expectations and Operating System Trade-offs

A necessary dose of reality must accompany any conversation about switching to Linux: it remains ill-suited for certain workflows. Many users who rely on specific professional production software have no practical way to migrate because their required toolsets either lack native ports or fail to run reliably through compatibility layers. This restriction applies directly to Gamers Nexus's own primary video production workstations, which remain firmly tethered to Windows. However, for users whose daily routines revolve around web applications, general computing, and PC gaming, Linux has progressed to a point where it represents a genuinely viable, high-performance alternative.

For anyone curious about testing the waters, Steve recommends installing a gaming-centric distribution onto a secondary solid-state drive to experiment without destroying an existing Windows installation. This trial run offers an exit route from invasive platform policies. Steve expresses deep frustration with Windows pushing mandatory, always-online requirements, arguing that an operating system demanding permanent internet connectivity to function is fundamentally unacceptable. Although the Gamers Nexus team suffered crashes, hard freezes, anti-cheat roadblocks, and half-hour-long shader compilation delays throughout their Linux test suite, the overall state of gaming on the platform remains better than at any previous point in PC history.

> “Moving away as much as it can from bypasses to things like always online connectivity, which I personally think for an operating system is absolutely horrifying and should not be tolerated or accepted.” — Steve Burke

> “And if more users grit their teeth through some of the initial learning pains, we think that might mean that we can kind of escape from some of the Windows issues as a community and we'd count that as a win.” — Steve Burke

- Professional software barriers still keep production workstations bound to Windows, but gaming and web-first computing are in a far better state than in previous eras. - Test hurdles encountered during the process included anti-cheat failures, system freezes, outright application crashes, and shader compilation passes that stretched up to 30 minutes.

---

[05:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=327s) The Limits of Direct Windows-to-Linux Comparisons

A critical caveat for interpreting the benchmark figures is that readers must avoid cross-comparing these Linux results against Gamers Nexus’s historical Windows charts, except where specifically highlighted. Over the past decade, Gamers Nexus has developed a deeply refined, highly automated testing pipeline for Windows that automates test execution, data management, and automated image capture for visual quality checks. None of that custom automation exists for Linux yet. Refactoring their entire test harness for Linux will take significant engineering time, forcing this entire maiden round of Linux benchmarks to be executed and audited manually.

Beyond the manual execution, fundamental technical disparities prevent direct numerical cross-comparisons between the operating systems. Game binaries under test may run on distinct native APIs versus translation layers, game titles may differ across OS builds, and the underlying frame capture utilities measure frame rates and frame times using different capture points and mathematical methodologies. Furthermore, Linux distributions feature modular components that update far more rapidly than Windows, creating moving targets that can easily invalidate historical baselines if a testing suite drags on for too long.

```mermaid flowchart TD accTitle: Testing Pipeline Differences accDescr: Diagram illustrating why Windows and Linux gaming benchmarks cannot be linearly cross-compared.

subgraph WindowsPipeline["Windows Testing Pipeline"] W1["10-Year Automated Test Suite"] --> W2["Automated Screen & IQ Captures"] W2 --> W3["Windows Native Frame Capture Tools"] end

subgraph LinuxPipeline["Linux Initial Testing Pipeline"] L1["Fully Manual Test Execution"] --> L2["Manual Run Verification & Reruns"] L2 --> L3["Linux-Specific Capture Tools & APIs"] end

W3 -.->|"Incompatible Capture Math & APIs"| X["Do Not Cross-Compare Charts"] L3 -.->|"Different Native/Proton Game Builds"| X ```

> “You can't even cross compare because the software that captures the data itself is different.” — Steve Burke

> “Testing is sort of different from a real use and that you need to maintain a specific environment in order to do it properly for comparison.” — Steve Burke

- The existing decade-old Gamers Nexus automated benchmarking harness does not run on Linux, necessitating 100% manual test passes and vetting. - Linux capture utilities compute frame times and frame rate pacing using different metrics and hooks than Windows software, preventing linear comparisons between the two operating systems.

---

[07:29](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=449s) Selecting Bazzite as the Primary Test Distribution

To maintain scientific control, Gamers Nexus restricted the hardware evaluation to a single distribution rather than splitting focus across the fractured Linux ecosystem. Because graphics cards represent the most straightforward hardware component to isolate and measure, the team retained their standardized GPU testing rig and evaluated three primary distribution candidates: SteamOS, CachyOS, and Bazzite. While Valve’s SteamOS was the natural first preference given the team’s familiarity with the Steam Deck, the lack of an official desktop release eliminated it from consideration for standardized desktop GPU testing.

Between the remaining community favorites, CachyOS promised cutting-edge performance tweaks and kernel flexibility, but carried potential stability liabilities for long-term bench consistency. Bazzite ultimately won out because of its widespread community adoption, its explicit engineering focus on gaming out of the box, and overwhelming viewer requests. The distro's maintainers actively curate default software choices, which reduced arbitrary decision-making for the benchmarkers. Crucially, Wendell from Level1Techs strongly endorsed Bazzite as a dependable, highly stable platform for launching a standardized hardware testing initiative.

> “Thousands of you requested we test it specifically, and that's the most important part to us.” — Steve Burke

- **SteamOS**: The original first choice due to Steam Deck experience, but disqualified because it lacks an official desktop installer. - **CachyOS**: Considered for its bleeding-edge performance features and flexibility, but shelved due to possible stability trade-offs during controlled testing. - **Bazzite**: Selected as the final platform based on heavy community requests, gaming-centric defaults, maintainer accessibility, and Wendell's explicit recommendation for platform stability.

[08:36](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=516s) Why Bazzite Was Chosen: Immutability and Fair Testing

The primary reason for selecting Bazzite is that it is an immutable distribution, which marks a significant philosophical departure in system management compared to traditional Linux setups. In the enthusiast community, users frequently tweak, optimize, and min-max system settings or resolve issues manually as they crop up. However, hardware benchmarking carries fundamentally different requirements than everyday system administration. A testing platform must be strictly predictable, stable, and designed to eliminate subjective human judgment calls wherever possible. If testers were to optimize or debug an issue specifically for one vendor's hardware, fairness would require applying that identical troubleshooting effort to every other device on the bench—a workload that is unviable and fraught with risks of uneven expertise.

Relying on an immutable, out-of-the-box state effectively railroads the testing process into maintaining equal footing across all hardware vendors. Bazzite Desktop Edition is distributed in two main flavors: an Nvidia build and a non-Nvidia build—a split Steve humorously compares to "the S&P 500 at this point." Following advice from both the Bazzite developers and Wendell of Level1Techs, Gamers Nexus set up two dedicated drives (one for each image). Bazzite ships as an atomic bootable container image pre-stocked with drivers and launchers, eliminating post-installation configuration and replicating the default end-user experience while mitigating cross-contamination risks. While Bazzite is intentionally not designed to be the absolute highest-performance or most bleeding-edge operating system available, it drastically reduces friction for newcomers and avoids catastrophic disruptions from unexpected background changes.

```mermaid flowchart TD accTitle: Bazzite Distro Setup and Image Architecture accDescr: Diagram illustrating the dual-drive test setup splitting Nvidia and non-Nvidia hardware using Bazzite's atomic container images.

subgraph DistroSelection["Bazzite Desktop Edition (Atomic Images)"] direction TB NVDrive["Nvidia Drive Image (Preloaded Nvidia Drivers & Launchers)"] NonNVDrive["Non-Nvidia Drive Image (AMD / Intel Drivers & Launchers)"] end

subgraph BenchTesting["Test Platform Goals"] direction TB Immutability["Immutable System Core (Predictable & Recoverable)"] Fairness["Rigid Bench Defaults (Eliminates Vendor-Specific Tweaks)"] end

DistroSelection --> BenchTesting ```

Antheas, a developer on the Bazzite team, explained directly to Gamers Nexus why the project recently abandoned aggressive out-of-the-box system tweaks:

> “We used to include some gaming oriented kernel optimizations, but as of August, we do not anymore. We found that most of the testing for those changes involved running, for example, Rocket League at 600 Hz while doing something like compiling a kernel. Nobody does that or cares.”

Steve dryly notes that assuming nobody in the Linux community compiles a kernel while gaming is a “bold assumption,” jokingly anticipating the comments section. Antheas elaborated on the real-world harm caused by those earlier optimizations:

> “We found those optimizations could cause a 2 to 30% performance degradation where the higher end was on Intel CPUs with heterogeneous cores. This actually ended up hurting our reputation, which is very unfortunate because the version of one of the optimizations we used had a bug. Moving forward, we will no longer make these sorts of changes. So performance on Bazzite should mirror Fedora and Arch but with more stability as we skip minor versions and delay the major a bit.”

[11:37](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=697s) The Moving-Target Problem of Linux Version Control

Testing procedures on Windows benefit from predictable update behavior. During a typical one-week Windows test cycle, operating system updates are frozen. Historical cross-comparison and A/B testing over the course of a year confirm that Windows GPU data remains largely shelf-stable unless Microsoft issues an architectural overhaul, which tends to affect CPUs far more than GPUs. In Linux, however, the software stack consists of numerous independent layers that update asynchronously and at an aggressive pace. Testing is substantially slower on Linux because it lacks full automation and requires manual intervention, creating a serious risk of software drift.

The moving pieces include the base operating system (Bazzite), the games themselves (which frequently receive Linux-specific patches distinct from their Windows builds), the Steam client, the Proton translation layer, and the graphics drivers. If the test lab accidentally allows one unmonitored dependency to update while freezing others, a bug could slip in that breaks an isolated device without accurately representing normal user conditions. To manage this without modifying the core system, the team chose to freeze Bazzite, graphics drivers, and Proton for the duration of the test pass while keeping detailed logs of all software versions. In the future, using GOG releases might allow precise freezing of individual game builds. For now, the team acknowledges that benchmarking on Linux means accepting that data will become obsolete much faster and will frequently need to be scrapped and remeasured.

- Software stack variables tracked during testing: Bazzite OS release, game-specific patches, Steam client updates, Proton version, and GPU driver builds. - Methodology rule: Freeze the entire core OS, drivers, and Proton throughout the test window rather than updating piecemeal.

[13:21](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=801s) The Economics of Hardware Benchmarking Suites

Steve pulls back the curtain on the business reality of media benchmarking, explaining why Linux testing poses an economic challenge for a channel like Gamers Nexus. Running a full GPU suite across roughly 26 graphics cards with new validation routines, custom game passes, and statistical calibration produces a financial loss if it results in only a single standalone video. The economic viability of hardware reviews relies on data longevity: once an initial test baseline is established, new cards (such as an RTX 5090, 5080, or 5070) can be appended individually over the following months to amortize that labor cost.

For context, Steve revealed that the team logged 85 hours in December 2024 overhaul work alone just rebuilding the standard GPU benchmark suite. That total accounted for selecting new titles, calibrating test runs, validating 26 GPUs, developing error-metric pipelines, and crunching standard deviations—entirely separate from filming, editing, and scripting. That investment makes financial sense because Windows data remains dependable for at least a full quarter. On Linux, however, the platform shifts so rapidly that appending data incrementally over long periods is currently unreliable. As a result, comprehensive Linux GPU testing will likely be restricted to concentrated sweeps once or twice a year, capturing all numbers as quickly as possible within a frozen snapshot. Benchmarking this initial suite required three straight weeks of Mike’s dedicated lab time to navigate setup hurdles and software bugs. Even so, Steve emphasizes that the project is well worth pursuing given mounting user frustration with Microsoft: “It’s going to take us some time to get it optimized, but we hate Windows and Microsoft at the moment for the spyware. So, it’s worth doing. It’s worth the investment.”

[16:28](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=988s) Test Bench Qualifiers and Software Settings

To match the operational state used on Windows as closely as possible, Gamers Nexus configured Bazzite with explicit desktop settings: Adaptive Sync was set to "Never," screen tearing was enabled for full-screen windows to bypass artificial compositing caps, and the system power profile was set to "Performance." While advisors like Wendell cautioned about potential pitfalls like kernel spinlock performance degradation, CPU performance governors, and background scheduler quirks, the team elected to leave the OS at its plug-and-play defaults. If an issue surfaces under the standard distribution profile, that friction is considered a valid, authentic part of the platform's out-of-the-box user experience.

Steve acknowledges the heavy use of qualifiers throughout the methodology, noting with self-deprecating humor:

> “Is that enough qualifiers for you? Can you tell that we don't have a lot of experience publishing Linux benchmarks?”

Even though the underlying graphics suite and in-game quality presets mirror the Windows testing configuration wherever the settings exist, the team is deliberately limiting direct Linux-to-Windows chart comparisons. A core reason for this separation is that frame time and FPS data on Linux are captured through a completely different telemetry toolchain, relying on MangoHud rather than Windows-based capture hooks.

[17:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1047s) Benchmarking Pitfalls, Disclaimers, and Scope

Evaluating gaming performance across operating systems introduces fundamental capture and metric discrepancies. Measuring frame output on Linux via MangoHud versus utilizing PresentMon on Windows creates an inherently uneven comparison due to disparate software pipelines and low-level platform architecture. Steve Burke cautions viewers about direct cross-OS chart comparisons:

> “Our focus today is on numbers, not stability.”

Frame delivery mechanics differ notably on Linux; while the test suite intentionally permits tearing to capture uncapped performance, tearing support remains relatively recent in Wayland compositors. Furthermore, the testing scope excludes multi-hour soak tests to determine whether titles run indefinitely without crashing.

```mermaid flowchart TD accTitle: Linux vs Windows Capture Differences accDescr: Pipeline showing distinct capture stacks and display servers between platforms. subgraph Linux ["Linux Platform"] L1["Proton Layer"] --> L2["Wayland Compositor"] L2 --> L3["MangoHud Capture"] end subgraph Windows ["Windows Platform"] W1["DirectX / Vulkan API"] --> W2["DWM Desktop Window Manager"] W2 --> W3["PresentMon Capture"] end ```

Hardware selection is constrained to modern architectures, led by the Nvidia GeForce RTX 50 series and recent competitors, because broader retro-testing introduces frequent OS rebuild overhead. Hardware configurations feature an 8 GB frame buffer for the AMD Radeon RX 9060 XT and 16 GB on the RTX 5060 Ti, strictly reflecting lab availability rather than strategic bias.

- **Capture Tools:** MangoHud on Linux versus PresentMon on Windows. - **Tear Handling:** Screen tearing is explicitly enabled across runs, despite Wayland's evolving support for tearing updates. - **OS Rebuild Constraints:** Testing older GPU generations was deferred to avoid repeatedly wiping and reconfiguring platform images. - **VRAM Configurations:** Tested hardware includes the 8 GB variant of the RX 9060 XT and the 16 GB variant of the RTX 5060 Ti.

---

[19:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1167s) Dragon's Dogma 2 GPU Benchmarks (1080p, 1440p, 4K)

Testing the Windows release of *Dragon's Dogma 2* at 1080p Maximum settings utilized KDE Plasma 6.2.4, KDE Frameworks 6.11.0, Qt 6.8.2 on Wayland, and Proton 9.0-4, with earlier test sequences on the RX 9070 XT and RTX 5070 Ti run under Proton Experimental (bleeding-edge build from late September 2024). At 1080p, the RTX 5090 paces the field at 176 FPS average, though higher resolutions are required to decouple GPU throughput from CPU limits. AMD's RX 9070 XT challenges Nvidia's upper tier directly, averaging 140 FPS to lead the RTX 5080 (134 FPS) by 3.9% with parity in low percentiles. The RTX 5070 Ti struggles with erratic frame stability: individual run passes recorded 0.1% lows swinging from a severe dip of 23 FPS, up to 61 FPS, and reaching 96–97 FPS across final runs, causing substantial variance.

At 1440p, the RTX 5090 expands its advantage over the RTX 5080 to 37% (154 FPS average). The RX 9070 XT holds parity with the RTX 5080, maintaining strong value against the RTX 5070 Ti. Contextualized against retail pricing—where the RTX 5070 Ti commands $730 to $750 while the RX 9070 XT retails between $600 and $650—AMD claims an advantage in both raw throughput and price-to-performance. Further down the stack, the RX 9070 ($550–$570) outperforms the $500 RTX 5070 by 19%, while the RTX 5060 8 GB runs out of memory, completely collapsing into unplayable frame rates.

At 4K, the RTX 5090 stretches its gap over the RTX 5080 to 44%, while the RTX 5070 Ti equalizes with the RX 9070 XT. The RX 9070 trails its XT counterpart by roughly 10%, keeping ahead of the baseline RTX 5070. The entry-level RTX 5060 8 GB could not complete the 4K test run due to extreme resource starvation.

---

[21:49](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1309s) Black Myth: Wukong Frame Pacing Collapse on Nvidia

Rasterized testing in *Black Myth: Wukong* reveals an architectural frame delivery flaw unique to Nvidia on this software stack. While Windows performance in this title leans toward GeForce hardware, Linux execution yields broken frame pacing across the entire RTX 50 lineup. At 1080p, the flagship RTX 5090 outputs a superficially strong 115 FPS average, yet its 1% and 0.1% percentiles crash relative to the mean. Steve Burke highlights the severity of the issue:

> “Even the RX 9060 XT is a better experience than the 5090, as its lows are more proportional to the average.”

```mermaid flowchart LR accTitle: Nvidia Frame Pacing Anomaly in Black Myth Wukong accDescr: High frame delivery instability on RTX 5090 compared to flat delivery on RX 9070 XT. subgraph RTX_5090 ["RTX 5090 Frame Delivery"] A["8 ms Baseline"] -->|"Frequent 10-12 ms Excursions"| B["18-22 ms Spikes"] B --> C["Visible Microstutter"] end subgraph RX_9070_XT ["RX 9070 XT Frame Delivery"] D["10-11 ms Flat Baseline"] -->|"Minimal Deviation <= 3-4 ms"| E["Consistent Delivery"] E --> F["Smooth Visual Pacing"] end ```

The AMD Radeon RX 9070 XT trails the RTX 5090 by just 9% in raw average FPS, but delivers a markedly smoother experience. Frame pacing metrics clarify why the RTX 5090 underdelivers: its frametime intervals bounce erratically from an 8 ms baseline up to recurring 18–20 ms spikes, creating noticeable visual stuttering. In contrast, the RX 9070 XT exhibits almost flat frame-to-frame intervals, with transient variances never exceeding 3 to 4 ms. The problem spreads across the product family—affecting the RTX 5080, 5070 Ti, 5070, 5060 Ti, and 5060—pointing to an underlying driver-level defect within the Linux display pipeline. Even the budget-tier Intel Arc B580 demonstrates more consistent interval pacing than Nvidia's top cards.

- **RTX 5090 1080p Metric:** 115 FPS average, severely compromised by 1% and 0.1% percentile collapses. - **Hardware Frametime Variance:** RTX 5090 excursions jump 10 to 11 ms frame-to-frame (spiking from 8 ms to nearly 20 ms). - **Radeon Stability:** RX 9070 XT variance stays confined within a tight 3 to 4 ms delta relative to prior frames ($n-1$). - **Ubiquitous Vendor Issue:** Identical pacing spikes replicate across every RTX 50-series card evaluated. - **Alternative Hardware:** Both the RX 9060 XT and the Intel Arc B580 maintain cleaner frame proportionality than the flagship RTX 5090.

---

[24:29](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1469s) Black Myth: Wukong (1440p and 4K Performance)

At 1440p, the RTX 5090 increases its average frame rate buffer over the RX 9070 XT, but the severe interval oscillation persists. Frametime interval graphs demonstrate that the RTX 5090 shifts its baseline rendering latency to roughly 10–11 ms, while launching consecutive spikes up into the 19–22 ms range. These rapid 10–12 ms deviations cause perceptible microstutters during actual gameplay. Addressing the underlying pacing through driver updates would likely tame these erratic spikes, though evening out frame times would also lower top-line average FPS numbers. The RX 9070 holds an exceptional position here, effectively matching the RTX 5080 in average frame throughput while delivering far superior frame pacing and smoothness.

Resolving to 4K finally tips the balance back toward the RTX 5090 simply through brute rendering power:

> “So, we finally found a situation where the game load crushes the other cards enough that even an inefficient and poorly running 5090 can still outmatch them.”

At 4K, the sheer computational workload pins the pipeline, bringing the RTX 5090's percentile lows up primarily because the run-to-run variability stabilizes under heavy GPU saturation. The RX 9070 XT drops just behind the RTX 5080 in overall average frame rate at 4K, yet preserves tighter low-end percentile proportionality than the competing RTX 5080.

- **1440p Frame Latency:** RTX 5090 baseline rendering rests at 10–11 ms, jumping to 19–22 ms under recurring spikes. - **1440p Competitive Standing:** RX 9070 non-XT matches the RTX 5080 in raw frame rate while providing better frame pacing. - **4K Brute-Force Scaling:** The RTX 5090 establishes clear leadership at 4K, where massive rendering load absorbs the driver pipeline pacing inefficiencies. - **4K Hierarchy:** RX 9070 XT yields its average FPS edge to the RTX 5080, but maintains cleaner frame pacing consistency.

[26:36](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1596s) Resident Evil 4 GPU Benchmarks: 1080p to 4K

Gamers Nexus moves on to *Resident Evil 4*, noting that testing this title required pruning far more outlier passes than usual due to severe run-to-run consistency issues. At 1080p, the Nvidia GeForce RTX 5090 establishes a commanding lead, posting a 360 FPS average, which places it roughly 27% ahead of the RTX 5080. While higher resolutions are needed to fully eliminate CPU constraints, the 5090 clearly demonstrates that high-tier cards are beginning to pull away. Meanwhile, the AMD Radeon RX 9070 XT trails just behind the RTX 5080 across average, 1% low, and 0.1% low metrics, giving the 5080 a modest 7.8% advantage in average frame rate. Broadly speaking, this benchmark proves much less favorable to Nvidia on Linux than it typically is in a Windows environment; AMD appears disproportionately favored here, as an RX 9070 XT running that close to an RTX 5080 is an uncommon result. Further down the chart, the RX 9060 XT leads the Intel Arc B580 by about 48%, landing in roughly the same performance bracket as the RTX 5060. The 0.1% low standard deviation in this test is extraordinarily high overall, despite the average frame rates remaining consistent.

Moving up to 1440p, the RTX 5090 drops to a 316 FPS average, yet stretches its lead over the RTX 5080 to an impressive 51%—a dramatic climb over the 27% gap observed at 1080p, driven primarily by imposing a heavier GPU-bound workload. The RTX 5070 Ti and RX 9070 XT are essentially locked in a dead heat in averages, though the 5070 Ti logs lower 0.1% low metrics due to an inconsistent pass that GN elected to keep as representative real-world data. Intel's Arc B580 and B570 continue to struggle on raw throughput, trailing the RTX 5060 and anchoring the bottom of the stack. However, Steve Burke points out an unexpected bright spot for Intel: while the Battlemage GPUs post lower raw numbers, their frame pacing is remarkably solid, demonstrating some of the most stable, tightly grouped frame-to-frame intervals on the entire 1440p chart.

At 4K, the RTX 5090 pulls off a massive 66% lead over the RTX 5080, demonstrating how sheer brute force allows it to scale relentlessly as pixel pipelines fill. The RTX 5060 manages an acceptable average frame rate at 4K on paper, but its 1% and 0.1% lows crater into unplayable territory due to severe frame interval pacing breakdowns. The Intel B580 and B570 land near the bottom again, but both manage to outperform the RTX 5060's low metrics. In the middle tier, the RTX 5070 and RX 9070 remain separated by approximately 10 FPS in their averages.

- **1080p Metrics:** RTX 5090 averages 360 FPS (+27% vs. RTX 5080); RTX 5080 leads RX 9070 XT by 7.8% in average FPS; RX 9060 XT leads Intel Arc B580 by 48%. - **1440p Metrics:** RTX 5090 averages 316 FPS (+51% vs. RTX 5080); RTX 5070 Ti and RX 9070 XT tie in average throughput; Arc B580 and B570 sit at the bottom but maintain the best pacing consistency. - **4K Metrics:** RTX 5090 stretches to a 66% lead over RTX 5080; RTX 5060 suffers broken frame pacing; Arc B580 and B570 beat the RTX 5060 in percentile lows.

[28:38](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1718s) Starfield Benchmarks: Nvidia Driver Breakdowns and AMD Dominance

Opening the *Starfield* section with classic deadpan sarcasm, Steve remarks, “No one plays this game, so we figured we'd multiply 0 by zero and test it in Linux.” What unfolds on the charts, however, is a catastrophic showing for Team Green. The Radeon RX 9070 XT performs extraordinarily well, to the point where Nvidia GPUs are effectively non-functional from an architectural scaling perspective due to broken driver interactions with the game on Linux. Steve emphasizes that while an RTX 5090 is objectively superior hardware across every conceivable metric, the reality under Bazzite is that its performance falls completely behind both the RX 9070 and RX 9070 XT. The 5090 leads the 5080 by a meager 19%, the RTX 5070 Ti roughly matches the 5080, and the RTX 5070 trails right on their heels. “This is a clear indication that something is just broken in this chain of drivers somewhere and software, and it's falling on Nvidia,” Steve explains, adding that AMD has consistently captured a far wider competitive edge in Linux gaming compared to its usual standing on Windows.

At 1440p, the exact same upside-down hierarchy persists. The Radeon RX 9070 XT sits firmly atop the chart, shadowed immediately by the non-XT RX 9070. The RTX 5090 places third, followed by the RTX 5080, with the 5090 eking out a 24% lead over the 5080. While the Nvidia cards still stack logically against each other by name, their real-world numbers are completely suppressed relative to their silicon specifications. “Broadly speaking, Nvidia is getting wrecked here. It's not even close,” Steve notes.

When pushed to 4K, the RTX 5090 finally regains ground as raw compute load mounts, but the result remains far from impressive. The RX 9070 XT retains the lead with a 60 FPS average and expected low scaling, while the 5090 essentially ties it within run-to-run variance, flanked on the other side by the RX 9070. Steve likens Nvidia’s Linux struggle in *Starfield* to how Intel Arc GPUs historically felt under Windows: the silicon possesses immense compute capability, but remains trapped inside an external driver-imposed performance box. Notably, Intel Arc is missing entirely from these charts because the game outright refuses to run properly on Battlemage hardware under Linux.

- **1080p & 1440p Anomalies:** Both the RX 9070 XT and RX 9070 beat the RTX 5090 at lower resolutions; RTX 5090 leads RTX 5080 by only 19% (1080p) and 24% (1440p). - **4K Parity:** RX 9070 XT holds 60 FPS average, tying the RTX 5090; RTX 5090 fails to demonstrate its expected hardware advantage over AMD. - **Missing Hardware:** Intel Arc B-Series cards are absent from the test suite due to fundamental driver/game execution crashes.

[30:37](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1837s) Dying Light 2: 5070 Ti Pacing Issues and Resolution Scaling

Turning to *Dying Light 2*, the testing kicks off at 1080p rasterized where the RTX 5090 regains its familiar top spot, delivering a 43% average frame rate advantage over the RTX 5080. Yet, an analysis of the frametime metrics shows the RTX 5090 and 5080 running into an artificial performance ceiling; their 1% and 0.1% lows are constrained, reflecting an underlying driver or software boundary. Conversely, the RX 9070 XT and RX 9070 maintain 1% and 0.1% lows that track far closer to their overall averages. The RX 9070 XT ties the RTX 5070 Ti in average frame rate but beats it handily in low-end frame consistency. The Arc B580 delivers a respectable showing at 74 FPS average, proving quite playable and competing directly with the RX 9060 XT.

Steve pauses the benchmark breakdown to spotlight an egregious frametime pacing issue exclusive to the RTX 5070 Ti. Under standard GN lab operating procedures on Windows, extreme run-to-run variance is investigated by stripping out bad passes and rerunning them to isolate background OS interference; if the dips disappear, they are discarded as measurement noise. In Linux, however, the inconsistency on the 5070 Ti was fully repeatable across multiple passes and represented the card's authentic behavior. Across four separate test runs, frametime plots reveal pervasive instability, culminating in a severe, highly perceptible spike right around frame 3,000 where frametimes soared to roughly 46 milliseconds—an unmistakable visual hitch for the user.

```mermaid flowchart TD accTitle: RTX 5070 Ti Linux Run Consistency Flow accDescr: Diagram illustrating GN's decision tree when encountering severe frametime spikes on the RTX 5070 Ti.

subgraph Evaluation ["Test Variance Detection"] A["RTX 5070 Ti Benchmarked"] --> B["Detect Severe Run-to-Run Variance"] B --> C["46ms Frametime Spike Observed (Near Frame 3000)"] end

subgraph Methodology ["Standard vs Linux Handling"] C --> D{"Standard Windows Protocol"} D -->|"Historical GN SOP"| E["Strip bad runs & re-evaluate for background tasks"] C -->|"Linux Bazzite Protocol"| F["Execute 4 repeat passes"] end

subgraph Conclusion ["Final Determination"] F --> G["Spikes reproduce consistently across runs"] G --> H["Retain data on official charts as authentic Linux behavior"] end ```

Moving through 1440p and 4K, the standard hierarchy reasserts itself with distinct behavioral quirks:

- **1440p Results:** The RTX 5090 posts a 50% average FPS lead over the RTX 5080. The 5080 leads a dead heat between the RX 9070 XT and RTX 5070 Ti. The RX 9070 XT maintains cleaner frame pacing intervals than its Nvidia counterpart, while the Arc B570 averages 43 FPS. - **4K Demands:** The RTX 5090 holds 119 FPS average, but its lows remain disproportionately depressed compared to the 5080, pointing once again to an elusive software ceiling. - **Low-Tier Elimination:** The RX 9070 XT hits a 64 FPS average (tying the RTX 5070 Ti), while entry-level cards like the RX 9060 XT, RTX 5060, Arc B580, and Arc B570 collapse under 4K raster loads, remaining playable only as comparative scaling datapoints.

[32:50](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1970s) Cyberpunk 2077: Phantom Liberty Rasterization and VRAM Bottlenecks

In *Cyberpunk 2077: Phantom Liberty*, rasterized testing immediately reveals elevated variance across the board. The 0.1% low standard deviation hovers between 10 to 11 FPS across at least three separate GPUs, establishing this title as far more variable under Bazzite than in a controlled Windows environment. At 1080p, the RTX 5090 clocks in at a 195 FPS average, beating the RX 9070 XT by 32%. Despite this raw throughput advantage, the RX 9070 XT demonstrates vastly superior percentile lows compared to the 5090. The RTX 5090 only maintains a smooth visual experience through sheer brute-force compute: its frame drops, though large in absolute delta, remain high enough on the absolute scale that they do not manifest as jarring stutters. The RTX 5080 falls behind the RX 9070 XT in average FPS and falls drastically behind in low percentiles, while the RX 9070 roughly matches the 5080's average while sustaining better pacing. The Arc B580 puts up a solid 60 FPS average, trailing the RTX 5060 by 20% in averages while effectively matching it in 1% and 0.1% lows.

At 1440p, the RTX 5090 holds 148 FPS with 1% and 0.1% lows landing at 104 and 90 FPS respectively—a 46% to 50% lead over the RTX 5080 and RX 9070 XT, which perform identically in average throughput. Intel’s B580 falls to 40 FPS under the increased pixel workload. The RTX 5060 begins to suffer significantly due to the punishing combination of its narrow 8GB VRAM capacity and OS-specific driver overhead under Linux.

By the time the benchmark reaches 4K, low-VRAM architectures fall apart completely. While the RX 9070 XT struggles at a 48 FPS average and trails the RTX 5080 across all three metrics, the lower tier sees an unexpected reversal: the Arc B580, Arc B570, and RX 9060 XT all substantially outperform the RTX 5060, particularly in the 1% and 0.1% lows, where Nvidia's budget card completely exhausts memory bandwidth and capacity.

- **1080p Dynamics:** RTX 5090 hits 195 FPS (+32% vs. RX 9070 XT); RX 9070 XT achieves superior low percentiles over both RTX 5090 and RTX 5080. - **1440p Dynamics:** RTX 5090 maintains 148 FPS (+46–50% vs. RX 9070 XT and RTX 5080); RTX 5060 shows early signs of memory starvation. - **4K Collapse:** Severe memory pressure enables Intel Arc B580/B570 and AMD RX 9060 XT to decisively beat the RTX 5060 in low-end frame stability.

[34:45](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2085s) Baldur's Gate 3: Native Linux Build Testing and Nvidia Inconsistency

The testing shifts to *Baldur's Gate 3*, which Steve highlights as one of the most mechanically and architecturally complex titles in the entire test suite. Unlike pure Windows titles running through Proton layers, *Baldur's Gate 3* features both an official native Linux build and a standard Windows build, layered across Vulkan and DirectX 11 execution APIs.

GN focused initial efforts on the native Linux build running the Vulkan API at 1080p, but the results quickly proved wildly erratic for Nvidia hardware. Performance on GeForce cards proved completely sporadic and unreproducible across runs. In one of the most bizarre data anomalies of the entire benchmark suite, the flagship RTX 5090 charted below the mid-range RTX 5070 on the native Vulkan build. Addressing the obvious confusion viewers would have seeing such an inverted chart, Steve confirms that the team conducted exhaustive retests and evaluations, concluding: “We are well aware that the 5090 is below the 5070 and that obviously is not how these cards should behave, nor does it seem like a normal or good test result. But after retests and evaluation, the only thing consistent is inconsistency when it comes to this configuration.” The native Linux implementation under Vulkan presents deep structural execution flaws on Nvidia drivers, highlighting the unpredictable friction gamers still face when navigating native vs. translation layers on Linux.

[35:23](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2123s) Baldur's Gate 3 Linux Native & Windows Build Benchmarks

Testing the Linux native Vulkan build of *Baldur's Gate 3* proved to be an exercise in frustration, particularly on Nvidia hardware. The AMD Radeon RX 9070 XT managed a 238 FPS average at 1080p, but its low-percentile frame rates failed to scale proportionally, delivering an overall inconsistent experience compared to other titles. The RX 9060 XT slotted in predictably beneath it. Nvidia's results, however, were an erratic mess. While the test rig logged numbers, Gamers Nexus opted against putting error bars on the chart because the native implementation is outright broken on Nvidia GPUs. The exact same instability surfaced at 1440p, where Nvidia cards delivered completely unreliable figures and the RX 9070 XT, while yielding a passable average, exhibited poor consistency.

```mermaid flowchart TD accTitle: Baldur's Gate 3 Build Comparison on Bazzite accDescr: Comparing the broken Linux native Vulkan branch against the stable Windows DirectX 11 branch running through Proton. subgraph Native["Native Linux Build (Vulkan)"] direction TB N1["AMD Radeon GPUs"] -->|"Variable Lows, Passable Avg"| N2["RX 9070 XT: 238 FPS (1080p)"] N3["NVIDIA GPUs"] -->|"Completely Broken / Erratic"| N4["Unreliable Data (No Error Bars)"] end subgraph Proton["Windows Build via Proton (DirectX 11)"] direction TB P1["Top GPUs (RTX 5090, 5080, 5070)"] -->|"CPU/External Bottleneck"| P2["Capped near ~170 FPS across resolutions"] P3["Mid-Range (RX 9060 XT & RTX 5060)"] -->|"Normal Scaling"| P4["RX 9060 XT: 130 FPS Avg"] P5["Intel Arc (B580 & B570)"] -->|"Viable Performance"| P6["B580: 79 FPS Avg"] end ```

Switching over to the Windows DirectX 11 build of *Baldur's Gate 3* running through Proton on Bazzite restored behavioral sanity. At 1080p, the RTX 5090, RTX 5080, and RTX 5070 ran into a strict CPU-bound ceiling, with the RTX 5060 Ti brushing against that limit as well. The 8 GB RX 9060 XT delivered a 130 FPS average with percentile lows closely tracking the RTX 5060 8 GB, putting both cards substantially ahead of Intel's Battlemage flagship. The Intel Arc B580 posted an average of 79 FPS with respectable lows considering Intel's driver overhead, and the B570 scaled right in line with expectations from previous Windows testing.

At 1440p, the exact same hierarchy held with predictable downward scaling, as the upper trio of Nvidia cards remained pinned against platform bottlenecks. By 4K, the RTX 5090 actually posted a 170 FPS average—technically exceeding its 1440p result. While nonsensical in a GPU-constrained workload, it confirms that 1080p, 1440p, and 4K were all bound by external CPU or software constraints rather than graphics throughput. The RX 9070 XT hit 103 FPS average at 4K in the Windows DX11 path, trailing the 115 FPS average it managed under the native Vulkan build, though the two API implementations are not directly comparable.

[37:24](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2244s) Ray Tracing Benchmarks: Cyberpunk 2077, Black Myth: Wukong, and Dying Light 2

Ray tracing testing kicked off with *Cyberpunk 2077* at 1080p, where Nvidia maintained its traditional Windows performance superiority entirely through hardware silicon advantages. The RTX 5090 set the high mark with a 135 FPS average, leading the RTX 5080 by 46%. The RX 9070 XT landed at 72 FPS average, wedged between the RTX 5070 and RTX 5070 Ti. Intel's Arc B580 made a remarkably strong value showing by averaging 58 FPS to edge past the RTX 5060 Ti, while the baseline 8 GB RTX 5060 failed the test entirely and could not run.

At 1440p ray tracing, the workload hammered every card tested: - Only the RTX 5090 and RTX 5080 managed to hold averages above 60 FPS. - The RX 9070 XT crept toward the RTX 5070 Ti without matching it, failing to demonstrate any of AMD's earlier low-frametime advantages seen in rasterization. - Intel's B580 punched well above its price tier, drawing near the RTX 5070 while matching or slightly besting its percentile lows. - 4K proved unplayable across the board: the RTX 5090 managed 49 FPS with tight frame pacing, while the next-best RTX 5080 collapsed to 30 FPS.

In *Black Myth: Wukong* with Ray Tracing Medium and FSR Quality, Nvidia dominated while AMD collapsed. The RTX 5090 achieved a 113 FPS average (30% over the RTX 5080) with identical lows. AMD did not appear on the chart until below the RTX 5060, plagued by disproportionately terrible percentile lows that completely inverted the rasterization trends. *Dying Light 2* with ray tracing reinforced this dynamic: the RTX 5070 outperformed the RX 9070 XT outright, leaving only the RTX 5060 Ti and RTX 5060 beneath the non-XT RX 9070.

[39:28](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2368s) Realities of Linux Gaming: Nvidia Drivers, Anti-Cheat, and Broken Ports

Reflecting on the overall testing experience, Steve Burke observed that high-end hardware generally ran games on Bazzite without catastrophic roadblocks, though the rapid pace of open-source development means bugs can appear or vanish overnight. Low-VRAM cards like the RTX 5060 and Intel Arc GPUs ran into far more friction. Gamers Nexus also sought to clarify whether community hostility toward Nvidia on Linux stems from broken hardware or licensing philosophy. Citing Linux developer Antheas, Steve explained the transition:

> “Now with GSP, which is a co-processor on the card, they can shove all those fixes in the firmware of that co-processor, which is closed source, and distribute an open driver that communicates with that. It's just fancy software engineer lawyering.”

In practice, the Nvidia flavor of Bazzite ran just as smoothly as the AMD/Intel images on modern GPUs. Furthermore, Steve highlighted the refreshing openness of the ecosystem, noting that developers behind Bazzite and MangoHud reached out directly to collaborate, a stark contrast to Microsoft where contacts speak only under strict anonymity.

Despite the progress catalyzed by the Steam Deck, Steve cautioned that desktop Linux is not frictionless and outlined several glaring problem areas: - **Anti-Cheat Restrictions:** Kernel-level anti-cheat remains a massive wall; *F1 24* had to be excised from the test suite entirely because it refused to boot. - **Missing Features in RE Engine:** Neither *Resident Evil 4* nor *Dragon's Dogma 2* surfaced in-game toggles for ray tracing on Linux, requiring manual configuration tweaks that violated sterile testing protocols. - **Inferior Native Linux Ports:** Native builds often perform far worse than Windows binaries run through compatibility layers. *Total War: Warhammer III*'s native port lags behind Windows updates, forcing users onto Proton. *Baldur's Gate 3* dropped a native desktop Linux patch during testing that broke desktop installations because developer Larian overlooked non-Steam Deck setups, while also forcing borderless full-screen to the monitor's native resolution regardless of user settings. - **Engine Instability & VRAM Crashes:** In *Cyberpunk 2077*, enabling RT Medium caused the game to consistently crash on the second loading screen when reloading saves or backing out to the menu. Meanwhile, VRAM-limited GPUs failed catastrophically: the 8 GB RTX 5060 managed *Total War: Warhammer III* at 1080p, but crashed during the loading screen at every higher resolution. - **Intel Arc Driver Maturity:** Intel experienced the worst stability; on the B580, *Baldur's Gate 3* native would not launch, *Dragon's Dogma 2* crashed before benchmarks could complete, and *Starfield* could not load save files at all.

[44:05](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2645s) Hardware Bugs, MangoHud Logging Crashes, and Shader Caches

Testing hardware on Linux revealed acute stability problems, especially on non-flagship GPUs and Intel hardware. Gamers Nexus noticed desktop flickering and visual artifacts during general usage. Ray tracing support was heavily fractured: across their entire testing suite, Cyberpunk 2077 was the solitary title that even exposed an RT toggle for Intel Arc hardware. When testing stepped down to the Intel Arc B570, the card could not launch Cyberpunk 2077 with ray tracing enabled without crashing immediately. Cyberpunk 2077 with ray tracing active was not an Intel-exclusive problem, as it also crashed consistently on AMD's RX 7600 XT (referred to in the audio context as the 9600 XT / 7600 XT class), though Intel required the most troubleshooting concessions overall.

```mermaid flowchart TD accTitle: Linux Testing Issues accDescr: Pipeline showing hardware instability, software logging bugs, and shader compilation friction encountered during Linux benchmarking.

subgraph HardFaults["Hardware & Drivers"] Desktop["Desktop Artifacts & Flickering"] IntelRT["Intel Arc: RT missing in suite\n(Except Cyberpunk 2077)"] RTCrash["Cyberpunk RT Crashing\n(Intel B570 & AMD RX 7600 XT)"] end

subgraph LogIssues["MangoHud Logging Friction"] CrashExit["Games crash to desktop\nwhen logging completes"] Freezes["Minute-long freezes\n(Resolved by Alt-Tab)"] HeroicBreak["MangoHud broke on Heroic Launcher\n(freedesktop runtime EOL)"] end

subgraph SteamOps["Shader Overhead"] InitCompile["Initial Shader Compile:\nUp to 25 mins (NVIDIA / Starfield)"] DailyUpdates["Daily forced micro-updates\nfor Steam shader caches"] end

HardFaults --> LogIssues --> SteamOps ```

Data collection was further complicated by MangoHud, the software utility used for frame rate and frametime logging. The team ran into repeated bugs where games would abruptly crash straight to desktop the moment MangoHud concluded a logging run. Gamers Nexus contacted the MangoHud developer directly to discuss potential solutions, but for the duration of this testing phase, they simply had to work around the instabilities. Games would also occasionally freeze entirely for up to 60 seconds at a stretch, an issue bypassed only by alt-tabbing out to the desktop and back into the game window. The MangoHud on-screen overlay itself had to remain completely disabled during runs; earlier testing on the Steam Deck demonstrated that active overlays inject noticeable frametime stuttering and pacing anomalies.

Later in the benchmarking pipeline, MangoHud stopped functioning altogether within the Heroic Games Launcher, which was being used to launch the GOG release of Cyberpunk 2077. This raised serious concerns for testing consistency:

> “This was alarming mainly because it implied that something was updating outside of our controls despite being so explicitly careful and controlled with all the software versions.”

The failure appeared linked to freedesktop.org runtime policies that drop support for releases after a two-year lifecycle. The team managed to track down a temporary workaround via an obscure Reddit thread, only for the original poster to promptly delete the thread.

Outside of logging utilities, Steam’s shader caching system introduced immense waiting periods. The first-time shader compilation pass on Nvidia graphics cards stretched out to grueling lengths, taking up to 25 minutes for *Starfield* before the title could even reach the main menu. Furthermore, Steam consistently pushed daily micro-updates to shader caches across several titles, matching behavior long documented on the Steam Deck.

- Intel graphics cards lacked ray tracing options entirely in every suite game except Cyberpunk 2077. - The Intel Arc B570 crashed instantly when attempting to boot Cyberpunk 2077 with ray tracing enabled. - MangoHud runs frequently ended with immediate hard crashes to desktop upon stopping the benchmark capture. - Alt-tabbing out and back in was the only method found to recover from 60-second in-game freezes. - Nvidia cards experienced first-time shader compilations taking up to 25 minutes on titles like *Starfield*. - Continuous daily shader cache patches regularly altered game footprints, mimicking the Steam Deck experience.

---

[45:45](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2745s) Benchmarking Feasibility and Long-Term Testing Realities

Reflecting on the dataset, Gamers Nexus treated this project as an exploratory effort to determine whether Linux could fit into their established GPU review workflows. In Windows, established baseline data remains dependable enough that the team can drop a newly launched graphics card into existing charts one or two months after the primary data run without invalidating comparisons. On Linux, however, the underlying ecosystem evolves too rapidly for that workflow.

> “The biggest thing we've learned is that this is going to be difficult to get to the level we're at on Windows where we can just append a card in say a month or two months later after collecting the core data set... because it changes a lot more.”

The team contemplated alternative testing strategies, such as allowing rolling software and driver updates while cataloging every specific Proton and kernel version alongside the card entry on the charts. However, that approach was rejected because it undermines standard benchmarking integrity:

> “Now you're cross comparing new data to old data and if the percent difference swings as a result of that but you haven't retested the old stuff then it's not accounted for and it makes the new test look potentially unfairly good or bad performing depending on how the update went.”

To make regular Linux GPU benchmarking viable going forward, Gamers Nexus is evaluating alternative compromises. These include building a tailored suite exclusively out of titles known to behave well on Linux, rather than prioritizing parity with their Windows test library, as well as porting internal automation tools over to Linux to reduce manual operator hours.

The extensive time investment required by strict data-validation controls means Linux testing cannot yet happen at the cadence of standard Windows GPU reviews. Viewers looking to support the massive labor involved in Linux data collection and experimental testing were directed to the Gamers Nexus store, highlighting their retro I/O-themed coaster packs, the backer tiers for the upcoming *AI Dystopia* special reports, the vaporwave unhinged smiley t-shirt featuring a censored speech box, and the YouTube copyright takedown shirt cheekily dedicated to Bloomberg.

## Notable quotes > [03:25](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=205s) And if more users grit their teeth through some of the initial learning pains, we think that might mean that we can kind of escape from some of the Windows issues as a community and we'd count that as a win. — Steve Burke

## Mentioned Gamers Nexus, Steve Burke, Mike, Antheas, Wendell, Level1Techs, Microsoft, Windows, Linux, Bazzite, SteamOS, Valve, CachyOS, Fedora, Arch, Proton, Steam, Steam Deck, GOG, Heroic Games Launcher, freedesktop.org, Wayland, KDE Plasma, MangoHud, PresentMon, Vulkan, DirectX 11, Larian Studios, Intel, Intel Arc B580, Intel Arc B570, AMD, AMD Radeon RX 9070 XT, AMD Radeon RX 9070, AMD Radeon RX 9060 XT, AMD Radeon RX 7600 XT, AMD Ryzen 9800X3D, Nvidia, Nvidia GeForce RTX 5090, Nvidia GeForce RTX 5080, Nvidia GeForce RTX 5070 Ti, Nvidia GeForce RTX 5070, Nvidia GeForce RTX 5060 Ti, Nvidia GeForce RTX 5060, Dragon's Dogma 2, Black Myth: Wukong, Resident Evil 4, Rocket League, Starfield, Dying Light 2, Cyberpunk 2077: Phantom Liberty, Baldur's Gate 3, F1 24, Total War: Warhammer III

## Video description SUPPORT OUR WORK: RETRO I/O Coaster Packs! https://store.gamersnexus.net/products/gn-drink-retro-io-coaster-pack-4-custom-3d-coasters-100x100mm We're finally benchmarking GPU performance in Linux, first using the Bazzite OS following thousands of community requests specifically for this operating system. A lot of this is exploratory and research for establishing methodology, so we're still learning how to control the platforms and software for this benchmarking. In the very least, this allows us to start generating some exit velocity from Windows for some people. Linux still isn't for everyone. Some users, like our own production machines, are bound to Windows by compatibility requirements with certain software. But gaming has dramatically improved on Linux over the years and is developing fast (despite still having issues), and so it may slowly start to become more viable for gaming users in particular.

Learn about the special backer tiers with bonus videos and digital content: https://store.gamersnexus.net/ai-dystopia And our new Vaporwave Smiley T-shirt: https://store.gamersnexus.net/products/unhinged-vaporwave-smiley-t-shirt-cotton-black Or grab our sticker packs! https://store.gamersnexus.net/products/unhinged-ai-dystopia-computer-case-sticker-pack We also have a new REDACTED glass with nucleation base: https://store.gamersnexus.net/products/censored-unhinged-smiley-glass-nucleation-base And our Copyright Takedown T-Shirt: https://store.gamersnexus.net/products/copyright-strike-takedown-black-cotton-t-shirt

Watch our Linux benchmarking discussion with Wendell: https://www.youtube.com/watch?v=5O6tQYJSEMw Watch our Steam Machine benchmarks: https://www.youtube.com/watch?v=bWUxObt1efQ

Bazzite, if you're interested. We have no affiliation with the maintainers or distro beyond using it here: https://bazzite.gg/

Like our content? Please consider becoming our Patron to support us: http://www.patreon.com/gamersnexus

中文简报

RTX 50驱动出现问题，AMD在Linux游戏测试中超越Nvidia Gamers Nexus测试显示，由于严重的驱动和帧生成时间问题，AMD GPU在Linux环境下常常击败Nvidia，但操作系统的快速变动导致难以进行可靠的长期基准测试。 Gamers Nexus对运行在不可变Linux发行版Bazzite上的Nvidia、AMD和Intel GPU进行了全面的游戏性能评估。基准测试显示，在《星空》（Starfield）和《黑神话：悟空》（Black Myth: Wukong）等游戏中，AMD硬件往往拥有更平稳的帧生成时间，并出人意料地能与Nvidia旗舰产品平起平坐；而Nvidia GPU在Linux转译层下则频繁遭遇严重的帧时间异常波动、驱动不稳定以及软件瓶颈。Intel的Arc Battlemage显卡提供了尚可的中端稳定性和光线追踪性能，但在某些游戏中会完全无法启动。最终该团队得出结论，尽管Linux游戏对于普通用户而言已成熟为可行的Windows替代方案，但软件的快速变动、缺乏自动化工具以及频繁出现的性能倒退，使得进行长期的增量式GPU基准测试变得极其困难。 Gamers Nexus使用专注于游戏的不可变发行版Bazzite，对现代GPU在Linux上的表现进行了为期数周的基准测试评估，旨在检验其原始性能以及建立持续性Linux GPU测试套件的可行性。 (0:00) 在Linux上的多款游戏中，AMD Radeon GPU表现出明显高于Nvidia RTX 50系列显卡的帧时间一致性，由于驱动性能倒退，在1080p和1440p分辨率下的《星空》等游戏中甚至超越了旗舰显卡RTX 5090。 (28:38) 由于抓取挂钩（MangoHud对比PresentMon）、补丁版本以及转译层执行等方面的架构差异，Windows与Linux基准测试数据之间的直接比较在技术上是无效的。 (5:27) 内核、Proton更新、驱动程序以及每日着色器缓存微更新所带来的上游软件快速变动，使得在不进行定期全套重测的情况下，目前在Linux上进行长期的增量GPU图表追加是不可行的。 (45:45)

Reading summary

Gamers Nexus evaluated desktop Linux gaming performance on Bazzite across modern graphics cards, finding that while the platform is increasingly viable for gamers seeking to escape Windows data harvesting, significant driver flaws and software volatility persist. Across rasterized benchmarks, AMD hardware frequently demonstrated superior frame pacing over Nvidia, though software bugs, native port regressions, and rapid update cycles complicate testing. Bazzite Selection and Controlled Testing Constraints (8:36) Gamers Nexus selected Bazzite—an immutable, gaming-focused distribution—to test modern GPUs across two separate container images: one dedicated to Nvidia and the other to AMD and Intel. Bazzite's atomic architecture enforces consistent system defaults, preventing testers from applying subjective manual optimizations that would compromise benchmarking fairness across hardware vendors. To maintain baseline integrity, the team froze the core operating system, Proton versions, and GPU drivers during testing passes to avoid the software drift typical of Linux's modular update model. Because Gamers Nexus's decade-old automated testing harness cannot run on Linux, all benchmark runs had to be executed and audited manually. Steve Burke cautioned against cross-comparing these Linux results with Windows charts: frame time captures on Linux rely on MangoHud rather than Windows PresentMon hooks, Wayland compositors handle screen tearing differently, and translation layers introduce distinct rendering pipelines that invalidate direct numerical equivalence. Rasterization Pacing and Nvidia Driver Breakdowns (21:49) In pure rasterization, AMD hardware delivered competitive throughput and significantly more consistent frame delivery than Nvidia. In Starfield, driver-level software flaws on Linux severely crippled GeForce performance: both the RX 9070 XT and non-XT RX 9070 outperformed the flagship RTX 5090 at 1080p and 1440p, with the 5090 only managing a tie at 4K through brute computational force. In Black Myth: Wukong, every tested RTX 50-series card suffered from broken frame pacing at 1080p and 1440p, exhibiting frametime interval spikes from an 8–11 ms baseline up to 18–22 ms. In contrast, the RX 9070 XT delivered flat frame times within a 3–4 ms delta. Similar driver inconsistencies surfaced across the suite. The RTX 5070 Ti produced repeatable, severe 46 ms frametime spikes in Dying Light 2 that forced Gamers Nexus to log the stutter as authentic Linux behavior. While the RTX 5090 comfortably regained substantial leads at 4K in Resident Evil 4 and Cyberpunk 2077, AMD consistently held narrower performance gaps and tighter percentile low stability than is typical under Windows. Native Linux Port Flaws and Ray Tracing (35:23) Testing demonstrated that native Linux game builds can perform substantially worse than Windows binaries running through the Proton compatibility layer. In Baldur's Gate 3, the native Vulkan build produced erratic, non-reproducible frame rates on Nvidia hardware, with the RTX 5090 charting below the mid-range RTX 5070. Switching to the Windows DirectX 11 executable under Proton resolved these pacing anomalies and restored normal hardware scaling, despite hitting strict platform CPU limits near 170 FPS. Native ports also suffered maintenance gaps: Total War: Warhammer III lagged behind Windows game patches, while a patch for Baldur's Gate 3 briefly broke desktop installations because developer updates targeted the Steam Deck. When testing shifted to ray tracing in Cyberpunk 2077, Black Myth: Wukong, and Dying Light 2, Nvidia's hardware architectural advantages returned. In Cyberpunk 2077 at 1080p with ray tracing, the RTX 5090 led the RTX 5080 by 46% and the RX 9070 XT by 88%. AMD GPUs suffered steep percentile low drops under ray tracing, inverting their rasterized pacing advantages, while low-VRAM models like the 8 GB RTX 5060 failed the test or crashed entirely. Platform Stability Roadblocks and Review Economics (39:28) Gamers Nexus encountered acute system friction throughout testing. Kernel-level anti-cheat prevented F1 24 from launching entirely, and Starfield demanded up to 25 minutes of initial shader compilation on Nvidia cards. Data collection was hindered by MangoHud bugs that frequently crashed games to the desktop immediately after logging ended, alongside an unexpected runtime deprecation on Heroic Games Launcher. Intel Arc graphics faced the steepest hurdle, crashing upon booting Cyberpunk 2077 with ray tracing and failing to load save files in Starfield. These technical challenges complicate standard benchmarking economics. Unlike Windows, where frozen baselines permit outlets to append newly released GPUs into existing charts over several months, the rapid, asynchronous updates of Linux kernels, Proton layers, and drivers render historical data obsolete quickly. Gamers Nexus determined that incremental chart additions are currently unfeasible on Linux, restricting future evaluations to intensive, periodic sweeps. Gamers Nexus concludes that while Linux gaming via distributions like Bazzite provides a functional escape from Windows telemetry for web and gaming tasks, users must tolerate driver quirks, shader compile delays, and anti-cheat barriers. For media outlets, regular Linux hardware testing remains economically unsustainable until automated harnesses are ported and software dependencies reach greater stability.

精炼阅读摘要

Gamers Nexus 对基于 Bazzite 系统的 Linux 平台进行了为期数周的 GPU 游戏基准测试，指出 Linux 游戏生态虽已达到历史最佳可用状态，但在底层驱动适配、帧生成时间一致性以及光线追踪方面仍存在显著短板。测试强调，由于捕获底层逻辑与软件栈快速迭代的“动态目标”特性，Linux 跑分数据无法直接与历史 Windows 图表进行线性对比。系统选型与基准控制：为何采用不可变系统 Bazzite (8:36) 微软日益激进的数据遥测、强制在线要求以及系统臃肿化，促使大量 PC 玩家寻求替代方案。为确立可靠的硬件基准，Gamers Nexus 在评估官方桌面安装程序缺位的 SteamOS 以及稳定性存疑的 CachyOS 后，最终选定广受社区推荐的 Bazzite。Bazzite 具备开箱即用的游戏预设，并获得 Level1Techs 专家 Wendell 的稳定性背书。团队采用双硬盘方案分别隔离 Nvidia 与非 Nvidia 镜像，以杜绝跨硬件测试污染。跨系统对比限制：软件栈漂移与高昂测试成本 (5:27) 测试人员警告，不可将 Linux 数据与过往 Windows 图表混为一谈。Windows 平台依托成熟的十年自动化管线与 PresentMon 捕获，而 Linux 测试目前依赖完全手动操作与 MangoHud 工具，二者在时间戳捕获点和数学统计方法上存在根本差异。此外，Linux 由底层系统、Proton 兼容层、Steam 客户端、显卡驱动和游戏补丁高度模块化拼接而成，更新极为频繁，导致历史基准迅速失效。 Steve Burke 透露，仅 2024 年 12 月重建常规 Windows 显卡测试套件就耗费了 85 小时。Windows 数据的时效性通常可维持数月，允许后续逐步追加单张新显卡；但 Linux 环境各层组件异步更新，导致数据无法长期向后兼容，全面测试面临严重的边际收益与维护成本挑战。光栅化性能异动：Nvidia 帧平滑度崩溃与 AMD 逆袭 (21:49) 在传统光栅化游戏中，Nvidia 显卡遭遇了严重的驱动级帧生成时间（frametime）波动。在《黑神话：悟空》测试中，旗舰级 RTX 5090 虽取得 115 FPS 的平均帧率，但帧间隔在 8 毫秒基准与 18 至 22 毫秒的骤增之间剧烈跳动，出现明显的微卡顿；相比之下，AMD RX 9070 XT 的帧波动严格受控在 3 至 4 毫秒内，甚至千元级 Intel Arc B580 的帧平滑度都优于 Nvidia 顶级卡。在《星空》中，Nvidia 驱动更出现断崖式效能异常，导致 RTX 5090 的平均表现落后于 RX 9070 XT 与 RX 9070。在《消逝的光芒 2》中，RTX 5070 Ti 在连续四轮独立测试中均在第 3000 帧附近稳定复现 46 毫秒的严重掉帧。尽管 RTX 5090 在 4K 超高负载下凭借纯算力硬件规格拉开领先优势（例如在《龙之信条 2》4K 下领先 RTX 5080 达 44%），但在中低分辨率与特定标题下，Linux 驱动缺陷使 GeForce 系列的稳定性受到严重削弱。原生移植陷阱、光追现状与 8GB 显存崩溃 (34:45) 软件分发形态对运行表现有着决定性影响。《博德之门 3》的 Linux 原生 Vulkan 构建版本在 Nvidia 硬件上彻底失常，甚至出现 RTX 5090 跑分低于 RTX 5070 的反常倒挂；切换为通过 Proton 转译运行 Windows DirectX 11 版本后，各卡才恢复正常的平台瓶颈限制与线性阶梯。这表明 Linux 原生移植版因维护滞后，体验往往劣于转译层。开启光线追踪后，Nvidia 的硬件架构优势得以恢复。《赛博朋克 2077》1080p 光追测试中，RTX 5090 平均帧达 135 FPS，领先 RTX 5080 达 46%；而在《黑神话：悟空》中开启中等光追后，AMD 显卡的低百分位帧率全面崩溃。显存容量瓶颈在 Linux 环境下尤为致命：8GB 显存的 RTX 5060 在《龙之信条 2》4K 下直接无法运行，并在《全面战争：战锤 3》中高于 1080p 时因显存耗尽而在载入界面发生闪退。实际运行障碍：监测工具崩溃、着色器延迟与反作弊壁垒 (44:05) 基准测试流程中暴露出频繁的系统层摩擦。内核级反作弊系统依旧是无法逾越的障碍，《F1 24》等游戏完全无法在 Linux 下启动；Intel Arc 系列遭遇最严峻的软件成熟度问题，除《赛博朋克 2077》外多数游戏未提供光追开关，且多款大作发生启动闪退或无法读取存档。遥测工具 MangoHud 在记录结束瞬间会频繁诱发游戏直接闪退至桌面，且因依赖 freedesktop.org 运行时支持周期而一度在 Heroic 启动器中失效；部分游戏出现长达 60 秒的死锁，仅能通过切回桌面再切回画面来解除。此外，着色器预编译耗时冗长，RTX 显卡在《星空》初次启动编译耗时高达 25 分钟，且 Steam 频繁推送日常微更新，极大增加了隔离测试的控制难度。 Gamers Nexus 总结指出，对于纯粹以网页浏览和一般游戏为主的玩家，Bazzite 等 Linux 发行版已成为摆脱 Windows 强制监控的可行替代品；然而，专业生产力软件的缺失、内核反作弊壁垒、Nvidia 帧平滑度波动以及生态快速变动带来的维护代价，意味着 Linux 平台尚未准备好在常规硬件评测流程中完全取代 Windows。

Detailed notes — zh

## Background 本内容由知名硬件评测频道 Gamers Nexus 于 2025 年 11 月 25 日发布在 YouTube 平台。根据视频简介，该期专题响应了数千名社区用户的强烈诉求，首次基于面向游戏的不可变 Linux 发行版 Bazzite，开展了针对多款主流 GPU 性能的大规模探索性基准测试，旨在为希望摆脱 Windows 系统监控与劣化的玩家建立严谨的评测方法学。 ## TL;DR Gamers Nexus 历时数周攻关，基于 Bazzite 系统对 AMD、NVIDIA 和 Intel 的多款最新显卡进行了全面的游戏与光线追踪性能评测。在纯光栅化测试中，AMD 显卡（如 RX 9070 XT）展现出出色的能效比，且在《黑神话：悟空》与《星空》等游戏中，其帧生成时间稳定性远超遭遇驱动与系统兼容瓶颈的 NVIDIA RTX 50 全系；但在光线追踪场景下，NVIDIA 依然维持了硬件架构上的绝对统治力。与此同时，测试揭示了 Linux 游戏生态目前面临的严峻现实：不可用的反作弊软件、劣质的原生移植版本、Intel 显卡与小显存显卡的频繁崩溃，以及因驱动、Proton 和着色器高频变动导致硬件评测数据保质期极短、难以常态化扩展的方法论困境。 ## Atmosphere & style 全片贯穿着 Steve Burke 标志性的讽刺幽默与严谨求实的极客精神，开场便对微软激进的数据收集与系统劣化进行了辛辣批判。叙事风格硬核而坦率，面对反复出现的闪退、微卡顿以及反常跑分，团队毫无遮掩地展示了详尽的帧时间波动曲线和排错波折，既有攻克技术盲区时的兴奋与执着，也透露出面对 Linux 异步更新生态时高昂工时成本的商业无奈。整部视频宛如一场毫无滤镜的硬核实验室实录，带给观众专业、真实且充满说服力的第一手探索质感。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=0s) Gamers Nexus 耗时数周针对基于不可变发行版 Bazzite 的 Linux 游戏环境进行首轮 GPU 深度实测，旨在评估脱离 Windows 系统的可行性与各显卡的相对性能表现。 - [19:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1167s) 纯光栅化游戏测试中 AMD 显卡表现异常强劲，RX 9070 XT 屡次在 1080p 下反超或逼平 NVIDIA RTX 5080，且整体逐帧生成时间平滑度明显优于 NVIDIA 竞品。 - [21:49](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1309s) NVIDIA 显卡在《黑神话：悟空》等游戏中遭遇灾难级的微观帧时间振荡，但在开启光线追踪后仍旧维持对 AMD 的绝对统治力。 - [11:37](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=697s) 由于 Proton、底层驱动及着色器缓存高频更新，Linux 显卡测试数据缺乏长期“保质期”，导致新卡无法像 Windows 那样直接追加对比，测试成本极高。 ## Key takeaways - [08:36](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=516s) 测试选择 Bazzite 是因其属于开箱即用的不可变系统，能最大限度排除人为微调对不同显卡公平性的干扰。 - [08:36](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=516s) Bazzite 已彻底移除曾导致异构大小核 CPU（尤其是 Intel）性能下降 2% 至 30% 的激进游戏内核优化补丁，回归平稳策略。 - [17:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1047s) Linux 下的数据采集依赖 MangoHud，与 Windows 采用的 PresentMon 机制不同，两者绝对数据不具备直接跨平台横向对比的统计学意义。 - [19:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1167s) 在《龙之信条 2》1080p 测试中，RX 9070 XT 以 140 FPS 平均帧超越 RTX 5080（134 FPS），而 8GB 版 RTX 5060 在 1440p 下遭遇断崖式性能雪崩。 - [21:49](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1309s) 《黑神话：悟空》光栅化测试中，RTX 50 全系遭遇 Linux 驱动与系统级兼容问题，RTX 5090 帧时间在 8ms 与 20ms 间剧烈震荡导致极度顿挫，顺滑度甚至不及 Intel B580。 - [28:38](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1718s) 《星空》测试暴露出 NVIDIA 在 Linux 驱动链条上的严重故障，旗舰 RTX 5090 性能全面落后于 AMD RX 9070 XT 甚至非 XT 版本的 RX 9070。 - [34:45](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2085s) 《博德之门 3》的 Linux 原生 Vulkan 版本存在严重适配缺陷，RTX 5090 成绩甚至反常落后于 RTX 5070；改用 Proton 运行 Windows DX11 版本后数据方回归理性。 - [37:12](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2232s) 开启光线追踪后 NVIDIA 重现统治力，《赛博朋克 2077》1080p 下 RTX 5090 领先 RTX 5080 达 46%，而 AMD 显卡的光追性能与低点帧均遭遇断崖式溃败。 - [37:12](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2232s) Intel Arc B580 在光追及特定光栅化测试中展现出极佳的帧时间稳定性，并在《赛博朋克 2077》1080p 光追下反超 RTX 5060 Ti，凸显其性比价优势。 - [38:52](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2332s) NVIDIA 现代所谓的 Linux“开源驱动”仅限于外层 MIT 许可的内核模块，核心调度逻辑与补丁均被封包在 GSP 协处理器的闭源固件中。 - [41:13](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2473s) 反作弊系统不兼容（如《F1 24》直接无法进入）与劣质的 Linux 官方原生移植版本，依然是普通玩家从 Windows 迁徙至 Linux 的核心障碍。 - [44:05](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2645s) 测试工具链自身存在缺陷，MangoHud 在记录停止时偶发闪退，叠加 NVIDIA 显卡在《星空》下长达 25 分钟的着色器编译，使 Linux 测试极其耗时且脆弱。 - [45:45](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2745s) 因为 Linux 无法像 Windows 一样保持数据长期不漂移，任何新硬件加入理论上都需要重新全量测试所有旧卡，导致常规化评测在商业上难以维系。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=0s) 微软的商业模式与向 Linux 逃逸的动力 Gamers Nexus 主持人 Steve 带着一贯的讽刺幽默开场，直言微软建立了一个“令人印象深刻”的商业模式：本质上就是间谍行为，即数据收集与买卖。然而，正是微软对 Windows 系统日益严重的劣化（enshittification）以及越来越激进的数据监控，让以往看似高不可攀的 Linux 门槛变得不再那么难以逾越。随着 SteamOS 的崛起以及 Windows 体验的持续下滑，社区对 Linux 游戏测试的需求已不再局限于少数狂热的技术极客。为此，GN 团队投入了大量精力：历时整整三周的不间断基准测试，加上此前两到三周用于平台搭建、前期调研以及所谓的“测试台卫生”（bench hygiene，即如何在长期测试中维持系统的纯净与受控），终于拿出了第一批 Linux GPU 测试数据。首轮测试不仅是为了向社区提供脱离 Windows 的数据参考，也是 GN 团队自身的摸索与训练过程。期间他们遇到了无数崩溃与卡死，但都硬生生地趟了过来。Steve 提到，自己上一次将 Linux 作为主力日常系统（daily drive）还要追溯到 2009 年至 2012 年左右，此后大多仅在 NAS 等设备上接触。在筹备期间，他们还直接与 Bazzite 的维护人员取得联系，以确定更严谨的测试控制规范。 - 本期项目由观众通过 GN 官方周边商城直接赞助，新品包括团队在 Blender 中设计的复古 I/O 杯垫组（涵盖经典机箱软驱面板、复古配色手柄、IBM Model M 风格小键盘及 GN 旋钮机箱面板四个款式）。 - 同步发售的还有全新“失控蒸汽波笑脸”（Unhinged Vaporwave Smiley）T恤与周边贴纸包，该款设计也将出现在后续关于世嘉与任天堂开发机、卡带被突击搜查的英国调查专题纪录片中。 --- ### [03:25](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=205s) Linux 现状免责声明：它依然不适合所有人在展开具体跑分之前，Steve 给出了一剂“现实预防针”。Linux 现阶段绝对还没有达到适合每一个人的程度，这很大程度上取决于用户的具体工作流。某些依赖特定专业商业软件的用户，可能会发现除了 Windows 之外根本别无选择，因为他们使用的软件要么没有 Linux 原生版本，要么无法通过兼容层正常运行——GN 团队内部用于视频制作的生产力机器目前就面临这一现实限制。对于这部分日常被绑定在专业工作流上的系统，Windows 依然不可或缺；但如果日常使用主要集中在 Web 网页端并兼顾游戏，Linux 如今已经完全具备了可行性。 ```mermaid flowchart TD accTitle: Linux 适用人群决策流 accDescr: 根据用户的生产力工具和游戏需求，判断当前是否适合从 Windows 迁移到 Linux 的决策路径。 subgraph Evaluation["用户工作流评估"] A["评估个人核心需求"] --> B{"是否依赖特定 Windows 专业软件？"} end subgraph Decision["系统选择路线"] B -- "是" --> C["留守 Windows (生产力受限)"] B -- "否" --> D{"主要用途？"} D -- "重度 Web 办公 + 游戏" --> E["推荐尝试 Linux (如 Bazzite/CachyOS)"] D -- "观望体验" --> F["备用 SSD 试用 或组双系统"] end ``` 用户只需准备一块闲置的 SSD 装上 Linux 发行版插进机器，就能低成本试水；即使不合适，也可以随时无缝切回 Windows。针对微软越来越过分的强制联网要求，Steve 表达了强烈的反感：“我个人认为，对于一个操作系统而言，这绝对是令人毛骨悚然的，绝不应该被容忍或接受。”（“which I personally think for an operating system is absolutely horrifying and should not be tolerated or accepted.”）尽管在本次 Linux 测试中，团队经历了系统崩溃、画面冻结、反作弊冲突以及长达半小时的着色器编译等一系列痛点，但 Linux 现阶段的游戏兼容性确实处于历史最好水平。在测试方法学上，GN 提出了极其关键的声明： - **切勿直接横向对比 Windows 数据**：GN 在 Windows 平台上拥有高度自动化、打磨近十年的测试工作流（涵盖自动跑分、画面截图质量校验与数据归档），但在 Linux 上这套系统无法直接复用，本次全部转为人工手动测试。 - **环境差异过大**：Linux 与 Windows 的底层不同，游戏可能有原生与兼容层之分，API 也可能发生改变。 - **采集工具原理不同**：两个系统上用于抓取帧率与帧时间（frametime）的捕获软件机制存在差异，线性比对缺乏统计学意义。除非是视频中明确列出并经过严格控制的同屏对比，否则不应把本期图表中的 Linux 绝对帧数直接去套 GN 以往的 Windows 评测。 --- ### [07:29](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=449s) 为什么首发选择 Bazzite 发行版？在硬件测试中，GPU 是受外部变量干扰相对最可控的硬件类别，因此 GN 决定在标准测试平台上沿用其经典的 GPU 测试套件。对于初次涉足 Linux 游戏测试的发行版选择，团队在早期筛选出了三个主要竞争者：SteamOS、CachyOS 和 Bazzite。各发行版的权衡如下： | 发行版候选 | 优势 | 劣势 / 排除原因 | |:--- |:--- |:--- | | **SteamOS** | GN 团队在 Steam Deck 上已有丰富的使用与测试经验。 | 尚未发布针对普通桌面 PC 的官方正式版，因此首先被排除。 | | **CachyOS** | 极致激进，性能调校更前沿，系统灵活性高。 | 激进的更新节奏可能以牺牲测试稳定性为代价。 | | **Bazzite** | 专为游戏设计，拥有庞大的社区呼声；开发者做出的底层预设简化了测试配置；获得了 Level1Techs 专家 Wendell 的重点推荐。 | 相对 CachyOS 而言调校稍显保守，但稳定性更适合建立基准环境。 | 最终 Bazzite 脱颖而出，最重要的原因在于成千上万的观众在评论区明确要求测试该发行版，同时其开发团队也主动与 GN 取得联系，为规范测试流程提供了帮助。正如 Level1Techs 的 Wendell 在交流中所建议的那样，Bazzite 是一个成熟且稳定的起点，最适合用作 GN 建立 Linux 基准测试基准的试验场。 ### [08:36](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=516s) 为什么选择 Bazzite：不可变发行版与开箱即用的测试哲学对于首次系统性开展 Linux 显卡测试，Gamers Nexus 选择 Bazzite 的核心原因在于它是一个“不可变发行版”（Immutable Distro）。在 Linux 发烧友社区中，大家习惯于不断手动调校、优化、榨取极致性能（min-maxing）或是修复遇到的各种 Bug。但对于硬件测试而言，测试与日常使用有着截然不同的哲学：测试平台必须高度可预测、极其稳定，并且要最大限度排除测试人员的主观干预，避免测试结果因为人为的微调决策而产生偏差。如果为了某一张显卡的运行状况去单独做针对性优化或修补，那么出于公平性，就必须对参与测试的所有显卡投入同等强度的优化；团队既没有精力也没有足够的经验去完美做到这一点。因此，测试一套接近开箱即用状态的系统，反而能用一套极其严格的既定规则，将所有硬件厂商约束在公平的对比基准线上。 ```mermaid flowchart TD accTitle: Bazzite Dual Branch Setup accDescr: Diagram showing how Bazzite splits into Nvidia and Non-Nvidia branches on separate SSDs. subgraph S1["Bazzite 镜像体系"] A["Bazzite 桌面版 (原子化可启动容器镜像)"] --> B["NVIDIA 版本镜像 (预装专有驱动)"] A --> C["非 NVIDIA 版本镜像 (针对 AMD / Intel)"] end subgraph S2["测试平台部署"] B --> D["独立系统盘 A (跑 RTX 系列显卡)"] C --> E["独立系统盘 B (跑 Radeon / Arc 显卡)"] end ``` Bazzite 桌面版（Desktop Edition）主要分为两个版本：NVIDIA 版本与非 NVIDIA 版本。在咨询了 Bazzite 开发团队以及 Level1Techs 的 Wendell 后，团队为两个版本分别配置了完全独立的物理固态硬盘。Bazzite 本质上是一个原子化可启动容器镜像（Atomic bootable container image），系统预装了全部驱动程序和游戏启动器，安装后几乎不需要做任何额外配置，能够最大限度还原普通用户的默认使用体验。尽管它的设计目标是“稳定、兼容与可恢复性”，并非市面上性能最极致、最激进前沿的发行版，但它对新手的门槛极低，也避免了在基准测试过程中因某些底层特性悄悄损坏而引发翻车。关于性能调校，Bazzite 开发者 Antheas 向 Gamers Nexus 透露了一段内情：Bazzite 过去曾包含针对游戏调优的内核优化补丁，但截至当年 8 月已彻底移除。Antheas 解释道：“我们发现那些改动的测试场景大部分是在编译内核的同时以 600 Hz 运行《火箭联盟》（Rocket League）。根本没人会这么干，也没人在乎。”针对这番话，Steve 忍不住吐槽这是对 Linux 社区非常大胆的假设，并开玩笑表示正等待弹幕和评论区的反驳。Antheas 进一步补充解释道：“我们发现那些优化补丁会导致 2% 到 30% 的性能下降，其中降幅最大的是采用大小核异构设计的 Intel CPU。这甚至损害了我们的声誉，非常不幸，因为我们使用的某项优化版本存在 Bug。今后我们将不再进行此类改动。因此，Bazzite 的性能表现应该与 Fedora 和 Arch 相当，但由于我们跳过了小版本更新并适度推迟大版本发布，系统的稳定性会更高。” - **不可变设计优势**：底层系统只读且原子化更新，防止系统文件被误改导致测试环境漂移。 - **双驱动分支**：分别部署 NVIDIA 与非 NVIDIA 独立系统盘，杜绝跨架构驱动污染。 - **内核策略调整**：彻底剥离过去导致异构大小核（尤其是 Intel CPU）严重性能倒退的激进内核补丁，回归与上游 Fedora/Arch 一致但更平稳的节奏。 --- ### [11:37](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=697s) Linux 与 Windows 测试的方法论差异与版本控制难题在 Windows 平台上进行周期为一周左右的测试时，团队可以直接冻结系统更新。过去一整年跨版本的 A/B 对比与校准验证表明，Windows 显卡测试数据具有极高的“保质期”（Shelf Stability），除了偶尔出现的 CPU 调度重构等大改动外，显卡驱动和游戏数据的漂移非常微弱，暂停一周更新不会带来任何负面效应。然而在 Linux 上，技术栈的层级远比 Windows 复杂得多，各组件的迭代节奏极快且常常是异步的。加上目前 Linux 端缺乏完全自动化的测试脚本，整体测试节奏要缓慢得多。 Linux 堆栈中存在大量独立且高频更新的软件层：Bazzite 操作系统本身、游戏客户端本身（其获得的更新有时与 Windows 版本不同步）、包含关键功能迭代的 Linux 原生 Steam 客户端、核心翻译层 Proton 以及各大厂商的底层驱动程序。如果测试人员无意中遗漏并漏冻结了某一个组件，就可能引发特定硬件报错，而这种报错又无法代表普通终端用户的真实体验。团队目前的折中方案是在单次测试周期内严格冻结 Bazzite 系统镜像、底层驱动以及 Proton 版本，同时巨细靡遗地手工记录下所有已知组件的具体版本号。长远来看，团队正考虑改用 GOG 平台，以便能将游戏版本锁死在绝对固定的历史版本上。 --- ### [13:21](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=801s) 基准测试背后的商业现实与时间成本 Steve 难得地分享了硬件频道在商业运作层面的现实困境：如果不讲究方法，运行一整套完整的 GPU 基准测试是百分之百赔钱的。如果花费巨大人力把 26 张显卡在全新的游戏套件中跑一遍、进行全新的数据验证与校准，最后却仅仅产出一期视频（就像本次 Linux 专题所做的这样），这在商业上完全不可行。这类测试之所以能够维持运转，是因为 Windows 下的数据具备极佳的保质期（通常能维持至少一个季度），团队可以在后续单卡评测（例如 RTX 5090、5080、5070 相继发布时）中直接追加对比数据，通过长期的内容产出摊薄前期的研发与测试成本。在 2024 年 12 月，团队仅为了彻底翻新显卡测试套件就累计投入了 85 个工时，涵盖新游戏引入、基准线验证、26 张显卡的完整轮换、动画误差指标（Animation error metrics）以及标准差剔除等深层工作。这还不算后续的拍摄、剪辑与文案工时。相比之下，Linux 端极难保持数据长期不漂移，如果每次加入新硬件都必须把所有显卡从头到尾重测一遍，团队根本无法承受。因此现阶段的策略是每年仅集中进行一到两次密集型 Linux 显卡大测，并在极短时间内抢跑完所有显卡，以防系统更新导致环境变化。仅本次测试，测试员 Mike 就被整整消耗掉了三周的全职时间来跑分与排除故障。尽管代价高昂且测试流程尚未达到 Windows 那样成熟的水平，Steve 仍然明确表达了死磕到底的决心：“我们目前打心底讨厌 Windows 和微软的间谍流氓软件（Spyware）。因此做这件事是值得的，这项投资完全值得。” - **85 工时投入**：常规单次 Windows GPU 框架重构需要 85 小时的纯基准测试与统计学清洗工时。 - **数据追加模式**：Windows 数据具备季度级的保质期，可随时追加新显卡；Linux 目前尚无法做到这一点。 - **三周排错攻坚**：测试员 Mike 单独耗费三周工时才完成了本次跨架构 Linux 数据的初步采集。 --- ### [16:28](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=988s) 测试设置对齐、免责声明与性能捕捉工具为了尽可能与 Windows 端的标准测试环境平齐，团队在 Bazzite 的系统级图形设置中做出了对应调整：将自适应同步（Adaptive Sync）强制设置为“从不”（Never），允许全屏窗口画面撕裂（Screen Tearing），并将系统的电源模式统一设定为“高性能”（Performance）。针对 Wendell 等 Linux 专家提出的内核自旋锁性能损耗（Kernel spin lock performance loss）以及 CPU 调度策略调节器（CPU performance governor）等深度问题，团队并未进行手动深挖。这正是 Bazzite 这种开箱即用系统的意义所在——直接安装并运行，玩家在这个过程中会踩到的任何坑，本身就是评测该系统最真实的一部分。 Steve 直言不讳地承认团队在 Linux 跑分测试领域仍属于起步探索阶段，因此特意放低预期并设置了大量前置限定条件。本次测试尽可能沿用现有的 Windows GPU 游戏库与画面设置（部分 Linux 端缺失的特性除外），但团队大幅限制了跨系统（Linux vs. Windows）的直接数据并列对比，因为两者的帧率捕获管道存在根本性差异：在 Linux 环境下，团队采用 MangoHud 工具来记录底层帧生成时间数据，其统计方式与 Windows 端存在不可忽略的技术分歧。 > “我们目前打心底讨厌 Windows 和微软的间谍流氓软件。因此做这件事是值得的，这项投资完全值得。” —— Steve Burke ### [17:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1047s) 测试方法限制：工具差异、平台特性与版本控制在 Linux 环境下进行帧率与帧时间捕捉采用的是专用的 Linux 测试工具，而 Windows 下使用的则是 PresentMon。这种测试工具本身的分歧以及底层操作系统的巨大差异，使得直接跨平台对比犹如“拿苹果比橘子”（apples-to-oranges comparison）。虽然许多玩家极度渴望看到 Windows 与 Linux 的正面对决，但 Gamers Nexus 现阶段的核心关切非常纯粹：如果用户将 Linux 作为主力操作系统，各个 GPU 之间的相对性能表现到底如何，以便玩家能在不同显卡之间做出明智的购买抉择。此外，Linux 处理画面的机制与 Windows 存在显著不同；例如测试中团队一贯允许画面撕裂（screen tearing），但撕裂这一概念对 Wayland 显示协议而言仍然相对较新。本次测试并不包含长时间的系统稳定性验证。测试团队会忠实记录并汇报运行中遭遇的闪退或兼容性问题，但绝不会耗费五小时去验证《龙之信条 2》（Dragon's Dogma 2）在 Linux 下长时间连续运行是否稳定。测试重心完全落在“基准数据”本身，而非耐久稳定性。在平台搭建上，本次仅测试了部分最新的显卡型号（重点覆盖 RTX 50 系列等），因为在 Linux 下适配老架构显卡往往需要重新编译甚至重装整套操作系统构建版本，维护成本和时间黑洞难以承受。 > “I think a lot of people want them to be compared directly but really what we care about is just simply how do the GPUs perform if you're using Linux as your OS so that you can choose between one GPU and the other just like you do with Windows.” —— Gamers Nexus - **工具与平台限制**：Linux 捕捉工具与 Windows 平台的 PresentMon 原理不同，严禁将本期 Linux 数据与历史 Windows 数据进行跨平台直接拉通对比（渲染 API 与底层管线可能完全不同）。 - **显卡显存规格说明**：针对存在多显存版本的显卡型号，本轮测试中 RX 7600 XT（视频中表述为 960 XT/7600 XT 类产品）使用的是 8 GB 版本，而 RTX 4060 Ti / 5060 Ti 类产品使用的是 16 GB 版本；该配置差异完全是由测试期间多套测试流水线并行的显卡硬件现货调配情况决定的。 --- ### [19:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1167s) 《龙之信条 2》各分辨率实测数据测试首先运行了基于 Proton 转译的《龙之信条 2》（Dragon's Dogma 2）Windows 版本，分辨率为 1080p 最高画质。测试环境基于 KDE Plasma 6.4.4、KDE Frameworks 6.17.0、Qt 6.9.2、Wayland 以及 Proton 10.0-2E（其中 RX 7900 XT / 9070 XT 与 RTX 5070 Ti 属于较早批次，运行在 Proton Experimental 10.0-20250919 构建上）。在 1080p 下，RTX 5090 以 176 FPS 的平均帧率领跑全场；然而 RX 9070 XT 的表现让英伟达感到极其不安——其平均帧率高达 140 FPS，以 3.9% 的优势超越了 RTX 5080（134 FPS），且两者的低帧表现持平。反观 RTX 5070 Ti，其 0.1% Low 帧表现惨烈，单次测试中甚至测出过 23 FPS、61 FPS 以及两轮 96-97 FPS 的巨大波动，反映出极差的单次运行一致性。RX 9070 均帧与 RTX 5070 Ti 基本持平，但拥有更平稳的低帧，且平均帧领先 RTX 5070 达 20%。随着分辨率提升至 1440p，RTX 5090 达到 154 FPS，领先 RTX 5080 约 37%。RX 9070 XT 仍旧能与 RTX 5080 打得难解难分。结合当时的零售市场行情，RTX 5070 Ti 售价在 730 至 750 美元左右，而 RX 9070 XT 售价约在 600 至 650 美元；在此工况下，RX 9070 XT 实现了对 5070 Ti 在性能与价格上的双重击杀。与此同时，售价约 550 至 570 美元的 RX 9070，表现也全面压制了 500 美元的 RTX 5070（领先幅度达 19%）。8 GB 显存的入门卡 RTX 5060 在 1440p 下遭遇断崖式暴跌，游戏彻底失去可玩性。升至 4K 分辨率后，极高的硬件负载拉开了纯算力差距：RTX 5090 领先 RTX 5080 的幅度扩大至 44%，RTX 5070 Ti 与 RX 9070 XT 战平，而 8 GB 的 RTX 5060 则彻底无法完成 4K 测试。 > “The 970 XT is disturbingly close to Nvidia though, for Nvidia at least. Up at 140 FPS average, and leading the RTX 5080s 134 FPS result by 3.9%.” —— Gamers Nexus | 显卡型号 | 1080p 表现要点 | 1440p 表现要点 | 4K 表现要点 | |:--- |:--- |:--- |:--- | | **RTX 5090** | 176 FPS 领跑，需更高分辨率排除 CPU 瓶颈 | 154 FPS，领先 RTX 5080 达 37% | 极高负载下拉开身位，领先 5080 达 44% | | **RX 9070 XT** | 140 FPS，反超 RTX 5080 约 3.9% | 与 RTX 5080 处于同一性能档位 | 与 RTX 5070 Ti 基本持平，领先 9070 约 10% | | **RTX 5080** | 134 FPS，均帧略逊于 RX 9070 XT | 被 5090 拉开 37% 差距 | 均帧位列第二，但性价比受挫 | | **RTX 5070 Ti** | 均帧与 9070 相当，但 0.1% Low 产生 23~97 FPS 巨大方差 | 落后于 9070 XT，且售价高出约 100 美元 | 均帧追平 RX 9070 XT | | **RX 9070** | 均帧领先 RTX 5070 约 20%，低帧稳定 | 领先 RTX 5070 达 19%，性价比凸显 | 稳压 RTX 5070，紧追 9070 XT | | **RTX 5060 (8GB)**| 表现孱弱，落后于 16GB 竞品 | 性能发生雪崩式暴跌（坠崖），完全不可玩 | 显存溢出，完全无法运行测试 | --- ### [21:49](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1309s) 《黑神话：悟空》1080p：N 卡的帧生成时间灾难在 Windows 平台上，《黑神话：悟空》（Black Myth: Wukong）在开启光线追踪时大幅偏向英伟达，纯光栅化模式下虽然稍平衡，但英伟达依然占据明显优势。然而，当舞台切换到 Linux 下的 1080p 光栅化测试时，灾难发生了：RTX 5090 虽然跑出了 115 FPS 的平均帧率，但其 1% Low 和 0.1% Low 帧率发生了不成比例的雪崩，数值被彻底砸穿到了地表。这种极高的平均帧与极低的低帧之间的割裂，意味着实际游戏画面充满了顿挫感，体验极度恶劣。相比之下，RX 9070 XT 虽然平均帧率只有微弱劣势（RTX 5090 的均帧仅领先它 9%），但 AMD 的帧生成节奏（frame pacing）展现出了压倒性的平滑度。这并非旗舰卡的孤例，从 RTX 5090、RTX 5080、RTX 5070 Ti，一路到 RTX 5070、RTX 5060 Ti 及 RTX 5060，整条英伟达产品线全军覆没，表现出完全一致的病态帧生成模式。这显然是该游戏、Linux 英伟达驱动与操作系统底层组合产生的系统性缺陷。甚至连低阶的 RX 9060 XT 和英特尔锐炫 B580，其实际游戏体验和帧率一致性都远比顶级的 RTX 5090 更加舒适。 > “This is just an Nvidia problem with this combination of game, driver, and operating system. Even the RX 960 XT is a better experience than the 5090, as it's lows are more proportional to the average.” —— Gamers Nexus ```mermaid flowchart TD accTitle: Linux下黑神话悟空显卡帧生成状态对比 accDescr: 对比英伟达RTX 50全系在Linux下出现的帧生成时间锯齿震荡与AMD及Intel的稳定帧生成状态 subgraph NVIDIA_ISSUE["英伟达 RTX 50 全系 (驱动与系统冲突)"] A["RTX 5090 / 5080 / 5070 / 5060"] --> B["表面平均帧率较高 (如 5090 达 115 FPS)"] B --> C["帧生成时间在 8ms 与 20ms 间高频跳跃"] C --> D["1% 与 0.1% Low 帧暴跌，肉眼可见频繁顿挫"] end subgraph STABLE_PERF["AMD 与 Intel 正常阵营"] E["RX 9070 XT / 9070 / Intel B580"] --> F["平均帧略逊或持平"] F --> G["帧时间偏差严格控制在 3-4ms 以内"] G --> H["Low 帧与均帧高度成比例，体感极其丝滑"] end ``` --- ### [23:35](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1415s) 帧间隔曲线深度剖析与 1440p/4K 表现通过引入帧与帧之间的时间间隔图（Frame-to-Frame Intervals，即微观帧时间曲线）可以清晰看到病灶所在：RTX 5090 的典型渲染帧时间原本基准在 8 毫秒左右，但其波形图上呈现出极其密集且高频的断崖式脉冲，频繁激增至 18 到 20 毫秒。这种帧与帧之间高达 10 到 11 毫秒的剧烈振荡（Excursions）以极快频率连续发生，转化为人眼的直观感受就是微卡顿和撕裂感，对于售价昂贵的顶级旗舰而言完全不可接受。反观 RX 9070 XT 的微观波形曲线平整如镜，帧间波动极小，偶发的微小扰动也不会超过上下一帧 3 到 4 毫秒。 ``` RTX 5090 帧时间剧烈震荡形态 (毫秒): 20ms | /\ /\ /\ /\ (频繁产生 10-12ms 瞬时卡顿) | / \ / \ / \ / \ 8ms |__/ \__/ \__/ \__/ \___ (典型基准线) -----------------------------------> 时间轴 RX 9070 XT 帧时间平稳形态 (毫秒): 12ms |~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ (波动控制在 3-4ms 以内，极其顺滑) -----------------------------------> 时间轴 ``` 在 1440p 分辨率下，RTX 5090 虽然凭借暴力规格拉大了平均帧的领先优势，但帧时间病态依旧，基准时间上升到 10-11 毫秒的同时，尖峰脉冲同步放大至 19-22 毫秒；RX 9070 XT 无论在均帧比例还是实际游玩流畅度上依然全方位胜出。直到分辨率拉升至 4K，沉重的渲染负载终于压垮了其他中低端显卡，RTX 5090 凭借庞大的硬件规模在平均帧上建立了绝对统治，但这仅仅是因为“即便运行效率低下且节奏糟糕，其绝对算力依然能强行碾压对手”；其低帧数据在多次循环测试中依旧展现出极不稳定的离散状态。 > “The rapid succession of these spikes is what makes the experience bad for the player. At least in terms of how it actually feels to play the game.” —— Gamers Nexus - **微观帧时间振荡数据**：1080p 下 RTX 5090 帧时间在 8ms 基准上频繁突波至 18~20ms（偏差 10~11ms）；1440p 下基准上升至 10~11ms，突波攀升至 19~22ms（偏差 10~12ms）。 - **英特尔锐炫 B580 异军突起**：考虑到其入门级定价，Intel B580 在 Linux 版本的《黑神话：悟空》中帧率平滑度令人惊喜，表现大幅优于同价位段受驱动卡顿困扰的英伟达入门卡。 - **4K 极限压制逻辑**：4K 场景下 RTX 5090 的胜出并非驱动或 Linux 适配良好，纯粹是渲染压力巨大导致其余显卡撞上性能墙，5090 依靠过剩的原始光栅吞吐量实现的粗暴压制。 ### [26:36](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1596s) 《生化危机4》GPU基准测试（1080p、1440p、4K）在《生化危机4》（Resident Evil 4）的测试中，由于单次运行之间的稳定性问题，测试团队不得不剔除比其他游戏更多的异常轮次。在 1080p 分辨率下，RTX 5090 以 360 FPS 的平均帧率领跑，领先 RTX 5080 约 27%。此时显然需要更高的分辨率来摆脱 CPU 瓶颈，不过 RTX 5080 与 RX 9070 XT 已经明显没有受到 CPU 限制。RX 9070 XT 在三项核心指标（平均帧率、1% 低帧、0.1% 低帧）上均紧咬 RTX 5080，使 RTX 5080 仅保有 7.8% 的平均帧率优势。GN 指出，从大局来看，这种情况在 Linux 环境下对 NVIDIA 的利好远不如在 Windows 下明显：“AMD seems disproportionately favored here as compared to typically to be that close to a 5080 is not a good thing for Nvidia.”（与平常相比，AMD 在这里似乎受到了不成比例的青睐；能咬得离 5080 这么近，对 NVIDIA 而言绝不是件好事）。此外，RX 9060 XT 领先 Intel Arc B580 约 48%，整体性能与 RTX 5060 相当。本项测试整体的 0.1% 低帧标准差极高，尽管平均帧率在运行之间表现得出奇一致。升至 1440p 后，GPU 负载显著增加，RTX 5090 的平均帧率降至 316 FPS，但对 RTX 5080 的领先优势一举拉大到 51%，远超 1080p 时的 27%。RTX 5070 Ti 与 RX 9070 XT 打成平手，但 5070 Ti 因为其中一轮测试表现偏低，导致 0.1% 低帧成绩更差。GN 强调这是代表真实玩家体验的实测数据，因此予以完整保留。Intel 的 B580 和 B570 表现依然不佳，落后于 RTX 5060 并垫底，但鉴于其官方建议零售价远低于同场其他显卡，这也在意料之中；值得赞许的是，Intel 显卡展现出了全图表中最一致的逐帧生成时间（frametime pacing）。在 4K 分辨率下，RTX 5090 凭借强大的硬件规模直接进行“暴力碾压”，对 RTX 5080 确立了惊人的 66% 领先优势。RTX 5060 虽然平均帧率勉强过得去，但其 1% 和 0.1% 低帧成绩堪称灾难，暴露出严重的帧生成间隔抖动问题。Intel B580 和 B570 虽然依旧排在图表末尾，但表现均优于 RTX 5060。图表中游位置上，RTX 5070 与 RX 9070 的平均帧率差距约为 10 FPS。 - 1080p：RTX 5090 达到 360 FPS 平均帧率，领先 RTX 5080 约 27%；RTX 5080 仅领先 RX 9070 XT 7.8%。 - 1080p：RX 9060 XT 领先 Intel B580 约 48%，性能逼近 RTX 5060。 - 1440p：RTX 5090 达到 316 FPS，领先 RTX 5080 达到 51%；RTX 5070 Ti 与 RX 9070 XT 平均帧率基本持平。 - 4K：RTX 5090 领先 RTX 5080 达 66%；RTX 5060 因帧生成时间严重不均导致低帧成绩垫底，不如 Intel B580 与 B570。 --- ### [28:38](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1718s) 《星空》Linux基准测试：NVIDIA遭遇驱动滑铁卢随后进入《星空》（Starfield）的测试。GN 调侃道，虽然现在根本没人玩这款游戏，但他们还是决定“0 乘以 0”，看看它在 Linux 下的表现。测试结果令人大跌眼镜：NVIDIA 在这款游戏中因驱动与游戏的交互问题彻底崩盘。GN 坦言，客观而言 RTX 5090 在几乎所有规格指标上都绝对强于 RX 9070 XT，但在 Linux 环境的实际测试中，RTX 5090 的实际表现不仅输给 RX 9070 XT，甚至连非 XT 的 RX 9070 都打不过。测试中，RTX 5090 仅领先 RTX 5080 约 19%，而 RTX 5070 Ti 与 RTX 5080 表现持平，RTX 5070 也紧随其后。GN 直言：“This is a clear indication that something is just broken in this chain of drivers somewhere and software uh and it's falling on Nvidia.”（这明确表明驱动和软件链路中的某个环节彻底坏掉了，而这个锅落在了 NVIDIA 头上）。相反，AMD 在该游戏中的平均帧与 1% 低帧都表现得极度稳定，在 Linux 平台的测试优势远超其在 Windows 平台的相对排位。 ```mermaid flowchart TD accTitle: Starfield Linux Performance Chain Breakdown accDescr: Diagram showing how software and driver issues degrade NVIDIA performance in Starfield under Linux compared to AMD. subgraph AMD_Stack["AMD 驱动与软件链路"] A1["Mesa / RADV 驱动"] --> A2["Proton 兼容层"] A2 --> A3["RX 9070 XT / 9070 表现优异 (帧率正常且稳定领先)"] end subgraph NVIDIA_Stack["NVIDIA 驱动与软件链路"] N1["NVIDIA Linux 专有驱动"] --> N2["Proton / 游戏指令调用异常"] N2 --> N3["异常硬件天花板 / 性能锁死 (RTX 5090 表现不如 RX 9070)"] end ``` 在 1440p 分辨率下，同样戏剧性的局面继续上演：RX 9070 XT 稳坐榜首，RX 9070 紧随其后，两张 AMD 显卡贴得极紧。之后才是 RTX 5090 和 RTX 5080。此时 5090 对 5080 的领先幅度微升至 24%，虽然比 1080p 略好，但绝对算不上正常。整体而言，NVIDIA 在这里被彻底按在地上打。尽管 NVIDIA 各显卡按命名高低排出了座次，但性能阶梯却与硬件实力严重脱节。到了 4K 分辨率，RTX 5090 的处境略有改善，但依然没能扭转局面。RX 9070 XT 以 60 FPS 的平均帧率保持领先，低帧数据也正常缩放；RTX 5090 基本只能和 RX 9070 XT 打成平手，两者低帧差距完全处于误差范围内，而 RX 9070 则紧咬在 5090 身后。尽管 4K 高负载让 5090 勉强追了上来，但与其庞大的规格完全不相称。GN 评论称这感觉就像 Intel 显卡在 Windows 下的窘境——硬件明明极具潜力，却被困在一堵看不见的软件高墙之内。至于 Intel Arc 显卡，则根本没有出现在图表中，因为游戏在 Intel 硬件上直接无法正常运行。 - 1080p 与 1440p：RX 9070 XT 与 RX 9070 击败包括 RTX 5090 在内的所有 NVIDIA 显卡。 - 1080p：RTX 5090 仅领先 RTX 5080 约 19%，性能阶梯严重受限。 - 1440p：RTX 5090 领先 RTX 5080 约 24%，仍未摆脱驱动异常限制。 - 4K：RX 9070 XT 达到 60 FPS 平均帧率，与 RTX 5090 打成平手；Intel 显卡因完全无法运行而缺席。 --- ### [30:37](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1837s) 《消逝的光芒2》测试与 RTX 5070 Ti 帧时间剧烈波动接下来是《消逝的光芒2》（Dying Light 2）测试。在 1080p 分辨率下，RTX 5090 在平均帧率上夺回领先，比 RTX 5080 高出约 43%。然而，RTX 5090 和 5080 的低帧表现再次暴露出它们在 Linux 下遭遇了驱动或软件上限的撞墙现象。相比之下，AMD 的 RX 9070 XT 和 RX 9070 的低帧数据更紧贴其平均帧率。RX 9070 XT 的平均帧率与 RTX 5070 Ti 相当，但在 1% 和 0.1% 低帧指标上明显胜出。Intel B580 跑出了 74 FPS 的平均帧率，证明其具备完全可玩的流畅度，并且足以抗衡 RX 9060 XT。在这项测试中，RTX 5070 Ti 出现了极高的跨轮次波动（run-to-run variance）。GN 解释了其测试流程：如果在常规测试中遇到类似问题，他们通常会剔除异常数据并重新跑几轮，如果无法复现就会归咎于 Windows 后台进程等偶发性干扰；但由于这是在 Linux 环境下的探索性测试，在核实数据后确认这不是孤立异常，而是该配置在该系统下的真实特征：“In this case, we're including the data because we're still new to Linux testing. This appears to be a real result, not some test outlier.”（在这种情况中我们保留了数据，因为我们对 Linux 测试还算初来乍到。这看起来是一个真实的结果，而不是测试中的孤立异常）。帧时间分布图显示，在四轮测试中，RTX 5070 Ti 的逐帧生成时间整体极不平稳，在接近第 3000 帧处甚至出现了一次飙升至 46 毫秒的严重卡顿，这种长达 46ms 的突刺在游玩中是肉眼立即可感的。在 1440p 分辨率下，RTX 5090 的平均帧率领先 RTX 5080 达 50%，而 RTX 5080 则领先于基本打平的 RX 9070 XT 与 RTX 5070 Ti。RX 9070 XT 与 RX 9070 依旧保持了更优秀的帧间隔稳定性。Intel B570 测得 43 FPS 平均帧率。升至 4K 分辨率后，RTX 5090 取得 119 FPS 的平均帧率，但其低帧成绩相比 RTX 5080 依旧不成比例地偏低，再次印证了高端 NVIDIA 显卡在 Linux 驱动层撞上了某种天花板。RX 9070 XT 以 64 FPS 的平均帧率与 RTX 5070 Ti 打平。至于 RX 9060 XT、RTX 5060 以及 Intel B580/B570，均已无法在原生 4K 下胜任该游戏，但数据依然被保留以供参考缩放趋势。 - 1080p：RTX 5090 平均帧率领先 RTX 5080 43%，但低帧比例失常。 - RTX 5070 Ti 在测试中出现严重的单帧耗时激增，峰值达到约 46ms。 - 1440p：RTX 5090 领先 RTX 5080 50%；RX 9070 XT 在帧生成平滑度上超越 RTX 5070 Ti。 - 4K：RTX 5090 达到 119 FPS；RX 9070 XT 达到 64 FPS，与 RTX 5070 Ti 打平。 --- ### [32:50](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1970s) 《赛博朋克2077：往日之影》光栅化基准测试测试项目随后切换至《赛博朋克2077：往日之影》（Cyberpunk 2077: Phantom Liberty）的纯光栅化测试。测试伊始，GN 团队就注意到了极高的标准差：0.1% 低帧的标准差达到了 10 到 11 FPS，是迄今为止所有图表中离散度最高的，至少有三张显卡落在此波动区间内，表明该游戏在 Linux 下的波动性显著大于 Windows。在 1080p 下，RTX 5090 测得 195 FPS 的平均帧率，领先 RX 9070 XT 约 32%。然而，RX 9070 XT 的低帧表现再次大幅反超 RTX 5090。所幸 RTX 5090 凭借其恐怖的绝对性能，即使遇到掉帧，其低帧数值依然足够高，因此在实际体验中卡顿并不容易被察觉，属于典型的“暴力破局”。RTX 5080 在平均帧率上落后于 RX 9070 XT，低帧差距更为悬殊；RX 9070 的平均帧率逼近 RTX 5080，低帧同样优于后者。Intel Arc B580 测得约 60 FPS 平均帧率，落后 RTX 5060 约 20%，但两者的 1% 与 0.1% 低帧基本持平。在 1440p 下，RTX 5090 达到 148 FPS 平均帧率，1% 和 0.1% 低帧分别为 104 FPS 和 90 FPS，领先 RX 9070 XT 与 RTX 5080 约 46% 至 50%。RTX 5080 与 RX 9070 XT 平均帧率基本打平。Intel B580 在 1440p 高压下跌至 40 FPS，排在 RTX 5060 之后。而 RTX 5060 则因显存容量限制叠加 NVIDIA 在该系统上的驱动问题而陷入挣扎。在 4K 分辨率下，RTX 5060 彻底溃败；而 RX 9070 XT 在 4K 下同样吃力，仅录得 48 FPS 平均帧率，在各项指标上均落后于 RTX 5080。令人意外的是，Intel B580、B570 以及 RX 9060 XT 在 4K 低帧指标上均大幅超越了显存吃紧的 RTX 5060。 - 1080p：RTX 5090 平均帧率为 195 FPS，领先 RX 9070 XT 32%，但低帧稳定性不如后者。 - 1080p：RX 9070 XT 平均帧率及低帧全面压制 RTX 5080。 - 1440p：RTX 5090 领先 RX 9070 XT 和 RTX 5080 约 46% 至 50%。 - 4K：RX 9070 XT 跌至 48 FPS，被 RTX 5080 反超；RTX 5060 因显存和驱动双重暴雷表现垫底，低帧落后于 Intel B580/B570。 --- ### [34:45](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2085s) 《博德之门3》测试的复杂性与 Linux 原生版异常最后进入《博德之门3》（Baldur's Gate 3）的测试环节。这款游戏的测试矩阵极为复杂：它不仅同时存在 Linux 原生版本（Linux native build）和通过 Proton 运行的 Windows 版本，还分别支持 Vulkan 和 DirectX 11 两种图形 API。测试首先在 1080p 分辨率、运行 Vulkan API 的 Linux 原生版本下展开。然而，测试结果完全陷入了混乱。NVIDIA 显卡在这一原生版本下的表现极不稳定且反常，甚至出现了旗舰级 RTX 5090 的成绩排在 RTX 5070 之后的离奇状况。面对如此颠覆常理的结果，GN 团队反复进行了多轮重测与验证，但结论十分无奈：“We are well aware that the 5090 is below the 5070 and that obviously is not how these cards should behave, nor does it seem like a normal or good test result. But after retests and evaluation, the only thing consistent is inconsistency when it [comes to this configuration]”（我们非常清楚 5090 测出来落后于 5070，这显然不是这些显卡该有的表现，看起来绝非正常或理想的测试结果。但在多次重测和评估后，面对这个配置，唯一保持一致的就只有它的‘不稳定性’）。这表明 Linux 原生游戏在面对特定驱动架构时，同样可能遭遇严重的适配灾难。 - 《博德之门3》具有 Linux 原生版与 Windows 转译版，并涉及 Vulkan 与 DX11 多重 API 变量。 - 在 1080p Linux 原生 Vulkan 环境下，NVIDIA 显卡出现严重异常，RTX 5090 成绩甚至落后于 RTX 5070。 - 团队多轮重测证实该异常具有确定性的不可预测性，展现出部分原生 Linux 移植版在兼容性上的缺陷。 ### [03:52](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=232s) 《博德之门3》Linux 原生版本测试：极其混乱与严重的稳定性崩溃在《博德之门3》（Baldur's Gate 3）Linux 原生版本的 Vulkan API 测试中，NVIDIA 显卡遭遇了极其严重且荒谬的兼容性问题。RX 9070 XT 在 1080p 下录得了 238 FPS 的平均帧率，但其低点表现（lows）与平均帧完全不成比例，整体帧平滑度远低于预期。RX 9060 XT 紧随其后排在正常区间，而 NVIDIA 系列显卡的测试结果则彻底陷入了混乱与不可靠之中。测试本身确实跑出了数据，但整个 Linux 原生构建版本存在极大的 Bug，数据稳定性严重崩盘，因此 Gamers Nexus 在该项图表中甚至无法标注误差线。这种情况在 1440p 分辨率下同样重现。NVIDIA 显卡的数据依然四处乱飞、完全不可信。Steve 毫不留情地评价道：“测试再次跑出了这些数字，这也是实际发生的现象，但它显然就是彻底坏掉了。”（“Now, the tests again produced these numbers and it is what happened, but it's clearly just broken.”）虽然 RX 9070 XT 缺乏应有的帧生成稳定性，但好歹能交出一个像样的平均帧率成绩。 - **1080p Vulkan Linux 原生版**：RX 9070 XT 平均 238 FPS，但低点表现不成比例；NVIDIA 全系数据混乱崩溃。 - **1440p Vulkan Linux 原生版**：NVIDIA 成绩依旧不可信，RX 9070 XT 帧时间波动偏大，但平均帧可用。 --- ### [04:06](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=246s) 《博德之门3》Windows 版本在 Bazzite 上的表现：回归理性与瓶颈效应放弃糟糕的原生版本，转而在 Bazzite 系统下通过 Proton 运行 Windows DX11 版本后，所有显卡的测试曲线立刻回归正常轨迹。在 1080p 下，RTX 5090、RTX 5080 和 RTX 5070 均遭遇了明显的 CPU 性能瓶颈；RTX 5060 Ti 同样受到 CPU 瓶颈限制，但程度略轻。8GB 显存的 RX 9060 XT 跑出了 130 FPS 的平均成绩，低点表现与 8GB 版 RTX 5060 相当，二者均大幅领先于 Intel Arc B580。考虑到 Intel 显卡当前在驱动和架构上的劣势，B580 取得 79 FPS 的平均帧且低点表现尚可，已经算是不错的成绩；B570 紧随其后，两张 Intel 显卡的阶梯差距完全符合此前在 Windows 系统下的表现。 1440p 呈现出完全相同的缩放比例，顶层显卡依然撞上了 CPU 或系统底层瓶颈。到了 4K 分辨率，RTX 5090 跑出了 170 FPS 的平均成绩，甚至在数值上略微反超了自身的 1440p 成绩——在纯粹的显卡瓶颈场景下这种逆转绝不可能发生，这直接证明从 1080p 到 4K，RTX 5090 始终受制于非 GPU 因素（包括系统调度或软件层限制）。对于 RTX 5090 而言，这三个分辨率在性能上实际是等同的。同时，RX 9070 XT 在 4K 下跑出了 103 FPS 的平均成绩，虽然无法直接跨 API 对比，但作为参考，其在 Linux 原生 Vulkan 下曾录得 115 FPS。 - **1080p Windows 版 (DX11)**：RTX 5090 / 5080 / 5070 集体遭遇 CPU 瓶颈；RX 9060 XT 8GB 达到 130 FPS 平均帧；Intel B580 录得 79 FPS，B570 符合预期。 - **4K 分辨率瓶颈反常**：RTX 5090 达到 170 FPS（因外部/系统瓶颈，4K 成绩反而略高于 1440p），RX 9070 XT 录得 103 FPS。 --- ### [04:12](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=252s) 光线追踪测试：《赛博朋克 2077》中的 NVIDIA 统治力与 Intel 惊喜开启光线追踪（Ray Tracing）后，测试基调发生了根本性变化。在《赛博朋克 2077》（Cyberpunk 2077）1080p 光追测试中，NVIDIA 延续了其在 Windows 平台的绝对优势。硬件架构上的代际领先在 Linux 下被完全复现。RTX 5090 维持了 135 FPS 的惊人平均帧率，领先 RTX 5080 达 46%；RTX 5070 Ti 紧随其后。AMD 阵营的 RX 9070 XT 仅取得 72 FPS 的平均表现，落在了 RTX 5070 与 RTX 5070 Ti 之间。最令人意外的是 Intel Arc B580，它以 58 FPS 的平均成绩力压 RTX 5060 Ti，在性价比维度打出了极其漂亮的一仗。与之相对的是 RTX 5060，在测试中彻底崩溃，无法完成运行。 1440p 光追对于这批显卡而言是极其沉重的负担。RTX 5080 是最后一张平均帧能够勉强维持在 60 FPS 以上的显卡；RX 9070 XT 试图贴近 RTX 5070 Ti 但未能成功。在光追环境下，AMD 失去了此前面向高帧率时的低点优势，帧平滑度完全无法抗衡 NVIDIA。相反，Intel B580 凭借其硬件价格展现出了超高的价值比，不仅性能逼近 RTX 5070，低点帧表现甚至旗鼓相当乃至略有超出。到了 4K 极限光追场景，全系显卡几乎全军覆没：RTX 5090 虽保持了极佳的帧间隔节奏（低点紧贴平均值），但平均帧也仅有 49 FPS；而排名第二的 RTX 5080 平均帧已经骤降至 30 FPS。 - **1080p RT**：RTX 5090 达 135 FPS（领先 RTX 5080 46%）；RX 9070 XT 平均 72 FPS；Intel B580 跑出 58 FPS 逆袭 RTX 5060 Ti；RTX 5060 崩溃无法运行。 - **1440p RT**：RTX 5080 刚刚压过 60 FPS 线；Intel 表现远超其身价预期；AMD 帧生成稳定性优势荡然无存。 - **4K RT**：负载过重，RTX 5090 仅 49 FPS，RTX 5080 跌至 30 FPS。 --- ### [04:23](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=263s) 光线追踪横评：《黑神话：悟空》与《消逝的光芒2》中 AMD 的溃败在《黑神话：悟空》（Black Myth: Wukong）开启中等光追与 FSR 质量档的测试中，RTX 5090 录得 113 FPS 平均帧，领先 RTX 5080 约 30%，二者低点表现基本打平。然而 AMD 在该项目中遭遇了溃败：AMD 的首张显卡甚至直到 RTX 5060 之后才勉强上榜，并且其 1% 低点帧极其恶劣。这与光追关闭时的态势形成了鲜明逆转——NVIDIA 在光追底层上大幅拉开距离，而 AMD 的低点帧和整体性能均不成比例地暴跌。RTX 5060 虽在平均帧上十分挣扎，但低点帧表现尚可，维持在了可用门槛附近。紧随其后的《消逝的光芒2》（Dying Light 2）光追测试同样成了 NVIDIA 的单方面碾压局。次级的 RTX 5070 成绩直接反超了 AMD 的旗舰级 RX 9070 XT；而在非 XT 版本的 RX 9070 身后，只有 RTX 5060 Ti 和 RTX 5060 垫底。整体而言，在高端硬件平台上，Bazzite 运行绝大多数游戏都没有致命障碍，但 Linux 游戏生态的更新速度瞬息万变，许多今天记录的 Bug 可能明天就会被补丁修复。然而，对于小显存显卡（如 RTX 5060）以及 Intel 显卡，翻车的概率明显呈指数级上升。 - **《黑神话：悟空》中等光追/FSR质量**：RTX 5090 平均 113 FPS（超 5080 约 30%）；AMD 显卡全线跌落至 RTX 5060 之下，低点帧极差。 - **《消逝的光芒2》光追**：RTX 5070 全面击败 RX 9070 XT，NVIDIA 架构在光追工作流中保持统治地位。 --- ### [04:32](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=272s) NVIDIA 在 Linux 上的“开源”争议与 GSP 协处理器真相在测试前，外界普遍充斥着“NVIDIA 对 Linux 支持极差”的言论。Gamers Nexus 深入探究后指出，这种差评究竟是因为显卡“真的不能用”，还是因为开源社区纯粹看不惯 NVIDIA 的技术路线与 Linux 自由精神相悖？资深极客 Anthas 对此给出了极具洞察力的技术拆解。 Anthas 解释称，外界常说的 NVIDIA 开源驱动，实际上仅指采用 MIT 许可证的 Linux 内核模块（kernel module），其核心驱动本体依然是闭源且专有的。过去在没有 GSP（GPU System Processor，GPU 系统处理器）协处理器的年代，所有的硬件调校（例如多达 1 万行的显示器 EDID 修复补丁）全部堆砌在驱动源码中，这些是 NVIDIA 绝不外泄的商业机密，因此驱动整体闭源，导致 Linux 社区极度不满并引发了内核授权的法律灰色地带。如今有了 GSP 协处理器，NVIDIA 顺理成章地将这些专属逻辑与补丁统统打包塞进了硬件协处理器的闭源固件里，外层只提供一个干净的开源通信驱动。Steve 引述 Anthas 的话一针见血地指出：“这无非是软件工程师精明的法律修辞把戏罢了。”（“It's just fancy software engineer lawyering.”） ```mermaid flowchart TD accTitle: NVIDIA Driver Architecture Evolution accDescr: Diagram contrasting legacy closed-source NVIDIA driver architecture with modern GSP firmware-based structure. subgraph Legacy["传统架构 (Pre-GSP)"] L_Kern["闭源内核模块"] -->|"包含大量专有代码\n(如 10,000 行 EDID 修复)"| L_HW["GPU 硬件"] L_Kern -.->|"社区指责\n版权法律灰色地带"| L_Com["Linux 开源社区"] end subgraph Modern["现代架构 (GSP 驱动模式)"] M_Mod["开源内核驱动 (MIT 协议)"] -->|"指令通信"| M_GSP["GSP 协处理器 (闭源固件)"] M_GSP -->|"硬件底层调度"| M_HW["现代 GPU 核心"] M_Mod -.->|"表面开源合规\n法律工程师技巧"| M_Com["Linux 发行版系统"] end ``` 在实际的 Bazzite 体验中，NVIDIA 定制版与非 NVIDIA 版在现代架构显卡上的游戏运行表现几乎毫无差别。Steve 还特别提到，本次测试最令人振奋的是开源社区的极速响应——Bazzite、MangoHud 等开发团队在 GN 提及项目后主动联系协助沟通，这与 Windows 平台下微软员工必须在“严格匿名”前提下才敢开口说话的压抑环境形成了天壤之别。 --- ### [04:13](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=253s) Linux 游戏生态的现实壁垒：反作弊、烂原生移植与显存黑洞尽管 Steam Deck 让 Proton 兼容层名声大噪，但 Linux 游戏体验绝非坦途，普通玩家必须降低心理预期。Gamers Nexus 详细罗列了在实际测试中遭遇的一系列严重问题： 1. **反作弊软件直接拒载**：部分反作弊系统在 Linux 下完全拒绝启动，进而导致游戏无法进入。测试团队因反作弊报错直接被迫剔除了《F1 24》。鉴于反作弊技术广泛存在于各大主流网游中，这依旧是横亘在 Linux 游戏普及道路上最坚固的绊脚石。 2. **Capcom RE 引擎的光追缺失**：在《生化危机4 重制版》和《龙之信条2》中，Linux 客户端完全隐藏了光线追踪选项。尽管社区可能有极客向的绕过方案，但为了维持标准测试环境，GN 只能跳过这两款游戏在 Linux 上的光追测试。 3. **“原生即劣质”的 Linux 构建版**：Steam 默认优先下载 Linux 原生版本，但某些开发商的原生移植水平极其糟糕。《全面战争：战锤3》（Total War: Warhammer III）的原生 Linux 版更新与性能严重滞后，逼得绝大多数玩家只能倒退回去用 Proton 跑 Windows 版本；《博德之门3》也是中途仓促上线 Linux 原生版，结果因拉瑞安工作室（Larian）疏忽了非 Steam Deck 的桌面级 Linux 玩家，导致大量崩溃，且原生版全屏模式会强制锁定显示器物理分辨率而无视游戏内设置。 4. **特定场景的规律性崩溃**：在《赛博朋克 2077》开启光追中等设置后，测试机在“第二次加载界面”会百分之百崩溃。玩家可以顺利载入一次存档，但只要重新读档或退回主菜单，游戏便立刻闪退；关闭光追后该现象消失。 5. **小显存与 Intel 显卡的双重灾难**：8GB 显存的 RTX 5060 在 1080p 下能勉强跑完《全面战争：战锤3》，但分辨率只要一提升，立刻卡死在基准测试的加载界面。至于 Intel Arc 显卡更是问题成堆：B580 运行《博德之门3》Linux 原生版直接无法启动、《龙之信条2》无法坚持到基准测试跑完、而《星空》（Starfield）干脆无法载入存档。 - **反作弊封杀**：《F1 24》因反作弊完全无法运行被移出测试阵容。 - **移植事故**：《全面战争：战锤3》与《博德之门3》的原生 Linux 构建问题连连，Proton 模拟版表现反超原生版。 - **硬件兼容性重灾区**：RTX 5060 遭遇显存瓶颈全屏卡死；Intel Arc B580 在多款 3A 游戏中遭遇无法读档或闪退。 ### [44:05](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2645s) 硬件光追与测试工具链的稳定性陷阱测试团队在 Linux（Bazzite）系统下遭遇了极为严重的兼容性与稳定性挑战。在桌面环境下，Intel 显卡出现了明显的画面闪烁和伪影（artifacts）。在光线追踪（RT）支持方面，整个测试套件中除了《赛博朋克 2077》（Cyberpunk 2077）之外，其余游戏均未向 Intel 显卡开放光追选项；而当降级到 Intel Arc B570 运行时，该卡在开启光追的情况下甚至无法启动《赛博朋克 2077》，直接发生崩溃。虽然 AMD Radeon RX 7600 XT（部分口误作 960 XT）在开启光追后同样频繁崩溃，但 Intel 在 Linux 平台上的特殊问题显然更加棘手。性能记录工具 MangoHud 自身也带来了连锁反应。最典型的症状是在 MangoHud 完成数据记录的瞬间，游戏会频繁闪退回桌面。团队已就此与 MangoHud 的开发者取得联系并构思解决方案。此外，游戏在测试中还会发生长达一分钟的无响应冻结，唯一的临时缓解手段是通过 Alt-Tab 切出桌面再切回。为了规避早期在 Steam Deck 上发现的“OSD 覆盖层导致严重卡顿”的已知问题，团队在整个测试过程中强制隐藏了 MangoHud 的前端显示。更令人头疼的是，在测试的中后期，MangoHud 突然完全无法挂载到用于启动《赛博朋克 2077》的 Heroic Games Launcher 上。这种突发故障令团队格外警惕，因为它意味着在团队极其严苛的版本锁死控制之外，有某些底层组件发生了静默更新（推测可能与 freedesktop 平台对运行满两年的版本执行 EOL 终止支持策略有关）。当时 Reddit 社区曾有一篇帖子指出了对应的修复方案，但在团队排查完毕后原作者便删帖了。着色器编译（Shader Compilation）同样是 Linux 游戏绕不开的时间黑洞。Steam 的首次着色器预编译耗时极长，在 NVIDIA 显卡上尤为夸张，例如运行《星空》（Starfield）时，单次编译耗时长达 25 分钟。不仅如此，许多 Steam 游戏每天都会推送体积微小的着色器缓存日常更新，这种在 Steam Deck 上司空见惯的现象极大地破坏了基准测试的环境一致性。 - **Intel 光追兼容性**：测试套件中仅《赛博朋克 2077》识别到 Intel 的光追功能，且 Arc B570 开启即崩溃。 - **AMD 崩溃表现**：RX 7600 XT 在《赛博朋克 2077》开启光追时同样呈现持续性崩溃。 - **MangoHud 记录 Bug**：日志记录停止瞬间游戏频繁跳出，并偶发长达 1 分钟的假死现象。 - **着色器编译延迟**：NVIDIA 显卡在《星空》下的首发编译耗时最高达 25 分钟，每日频繁推送着色器缓存更新。 ### [45:45](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=2745s) 结论：Linux 基准测试的方法论困局与未来展望 Steve Burke 总结指出，本次测试本质上是一次极具探索性质的“摸底行动”，其得出的最核心教训在于：想要在 Linux 上达到 Windows 那种成熟、可扩展的测试流水线是极度困难的。在 Windows 体系下，测试团队建立基准数据集后，可以在一个月或两个月后直接插上一张新显卡运行并追加到图表中；但在 Linux 平台上，底层系统组件的更新频率和变动幅度远超预期，长期保持基准环境的纯净与统一近乎不可能。对于后续测试，团队讨论过一种妥协方案：允许系统和组件保持滚动更新，并在图表上通过透明的标注进行说明（例如标注具体的 Proton 版本更迭），但 Steve 明确对这种折中表达了排斥：“Now, there's another compromise we haven't talked about here, which is because I don't really like it.” 如果在缺乏控制的滚动更新下强行对比新老数据，一旦性能发生百分比层面的剧烈波动，而老显卡又没有被全量重新测试，新卡的数据就会显得极不客观。未来若想加快 Linux 测试的落地，团队可能需要采取妥协手段，例如特意挑选在 Linux 下原生或兼容表现极佳的游戏来建立子测试集，或者将部分自动化流程移植到 Linux。最后，Steve 强调了团队坚持严苛方法论的原因：“you could do it quicker, but you know, we we really have rigid processes to try our best to maintain data integrity and for something we're less familiar with.” 这种严谨流程导致测试极为耗时，因此现阶段 Linux GPU 基准测试无法像 Windows 评测那样频繁发布。他呼吁希望支持这类重度硬核探索的观众前往官方商城购买复古 I/O 杯垫、针对 AI 专题的赞助档位，或是 Vaporwave 笑脸 T 恤，以帮助团队维系高成本的研究支出。 ```mermaid flowchart TD accTitle: Linux GPU 评测环境的不稳定性闭环 accDescr: 展示 Linux 评测由于驱动、Proton 及着色器频繁变动，导致无法像 Windows 那样直接追加新显卡数据的因果图。 subgraph PlatformInstability["Linux 滚动环境不确定性"] A["系统底层/Freedesktop 变动"] --> B["Proton 与驱动频繁更新"] B --> C["着色器缓存日常更新"] end subgraph DataIntegrity["数据一致性危机"] C --> D["老数据基准迅速失效"] D --> E{"能否直接插卡追加测试？"} E -- "强行追加" --> F["新老显卡数据失真，对比失去公平性"] E -- "严格控制" --> G["每逢新卡必须全量重测全部旧卡"] end G --> H["测试工时呈指数级膨胀，常规评测难以持续"] ``` ## Notable quotes > [13:21](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=801s) 我们目前打心底讨厌 Windows 和微软的间谍流氓软件。因此做这件事是值得的，这项投资完全值得。 — Steve Burke > [17:27](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1047s) what we care about is just simply how do the GPUs perform if you're using Linux as your OS so that you can choose between one GPU and the other just like you do with Windows. — Gamers Nexus > [21:49](https://www.youtube.com/watch?v=ovOx4_8ajZ8&t=1309s) Even the RX 960 XT is a better experience than the 5090, as it's lows are more proportional to the average. — Gamers Nexus ## Mentioned Gamers Nexus, Steve Burke, Bazzite, Windows, Microsoft, Linux, SteamOS, CachyOS, Fedora, Arch, Level1Techs, Wendell, Antheas, Proton, Steam, MangoHud, PresentMon, Wayland, KDE Plasma, NVIDIA, AMD, Intel, RTX 5090, RTX 5080, RTX 5070 Ti, RTX 5070, RTX 5060 Ti, RTX 5060, RX 9070 XT, RX 9070, RX 9060 XT, RX 7900 XT, RX 7600 XT, Intel Arc B580, Intel Arc B570, Dragon's Dogma 2, Black Myth: Wukong, Resident Evil 4, Starfield, Dying Light 2, Cyberpunk 2077: Phantom Liberty, Baldur's Gate 3, Total War: Warhammer III, F1 24, Heroic Games Launcher, GOG, Larian Studios, Capcom, Vulkan, DirectX 11 ## Video description SUPPORT OUR WORK: RETRO I/O Coaster Packs! https://store.gamersnexus.net/products/gn-drink-retro-io-coaster-pack-4-custom-3d-coasters-100x100mm We're finally benchmarking GPU performance in Linux, first using the Bazzite OS following thousands of community requests specifically for this operating system. A lot of this is exploratory and research for establishing methodology, so we're still learning how to control the platforms and software for this benchmarking. In the very least, this allows us to start generating some exit velocity from Windows for some people. Linux still isn't for everyone. Some users, like our own production machines, are bound to Windows by compatibility requirements with certain software. But gaming has dramatically improved on Linux over the years and is developing fast (despite still having issues), and so it may slowly start to become more viable for gaming users in particular. Learn about the special backer tiers with bonus videos and digital content: https://store.gamersnexus.net/ai-dystopia And our new Vaporwave Smiley T-shirt: https://store.gamersnexus.net/products/unhinged-vaporwave-smiley-t-shirt-cotton-black Or grab our sticker packs! https://store.gamersnexus.net/products/unhinged-ai-dystopia-computer-case-sticker-pack We also have a new REDACTED glass with nucleation base: https://store.gamersnexus.net/products/censored-unhinged-smiley-glass-nucleation-base And our Copyright Takedown T-Shirt: https://store.gamersnexus.net/products/copyright-strike-takedown-black-cotton-t-shirt Watch our Linux benchmarking discussion with Wendell: https://www.youtube.com/watch?v=5O6tQYJSEMw Watch our Steam Machine benchmarks: https://www.youtube.com/watch?v=bWUxObt1efQ Bazzite, if you're interested. We have no affiliation with the maintainers or distro beyond using it here: https://bazzite.gg/ Like our content? Please consider becoming our Patron to support us: http://www.patreon.com/gamersnexus RELATED PRODUCTS [Affiliate Links] AMD RX 9070 XT on Amazon: https://geni.us/BvWWM AMD RX 9070 on Amazon: https://geni.us/7Fgt NVIDIA RTX 5070 on Amazon: https://geni.us/ZUB8fT Intel B580 on Amazon: https://geni.us/V1GVXd NVIDIA RTX 5060 on Amazon: https://geni.us/kmtD6 AMD Ryzen 9800X3D CPU on Amazon: https://geni.us/bsS1kFv TIMESTAMPS 00:00 - Benchmarking Gaming on Linux 03:25 - It's Not For Everybody and Disclaimers 07:29 - Choosing a Linux Distribution 11:37 - METHODOLOGY: Testing Considerations 13:21 - METHODOLOGY: Some Business Context 16:28 - METHODOLOGY: Version Control and Qualifiers 19:27 - Dragon's Dogma 2 GPU Benchmarks 1080p 20:39 - Dragon's Dogma 2 GPU Benchmarks 1440p 21:25 - Dragon's Dogma 2 GPU Benchmarks 4K 21:49 - Black Myth Wukong: 1080p 23:35 - Black Myth Wukong Frametime Intervals 1080p 24:29 - Black Myth Wukong: 1440p Linux Benchmarks 25:31 - Black Myth Wukong Frametime Intervals 1440p 25:58 - Black Myth Wukong: 4K 26:36 - Resident Evil 4 1080p Benchmarks 27:28 - Resident Evil 4 1440p 28:07 - Resident Evil 4 4K 28:38 - Starfield 1080p 29:32 - Starfield 1440p 29:58 - Starfield 4K 30:37 - Dying Light 2 1080p 31:15 - Dying Light 2 5070 Ti Frametime Issue 32:01 - Dying Light 2 1440p 32:22 - Dying Light 2 4K 32:50 - Cyberpunk 2077 1080p 33:51 - Cyberpunk 2077 1440p 34:20 - Cyberpunk 2077 4K 34:45 - Baldur's Gate 3 Linux Native 1080p 35:51 - Baldur's Gate 3 Linux Native 1440p 36:06 - Baldur's Gate 3 Windows Build 1080p 36:41 - Baldur's Gate 3 Windows Build 1440p, 4K 37:24 - Ray Tracing: Cyberpunk 1080p 38:02 - Ray Tracing: Cyberpunk 1440p, 4K 38:39 - Ray Tracing: Black Myth, Dying Light 2 39:28 - PROBLEMS: Bugs, Anti-Cheat Software, Shader Cache 45:45 - Conclusion CORRECTIONS: 19:27 - The subtitle for the charts says "Win11." These are done on Bazzite, which is in the chart major title (correctly). The subtitle incorrectly states Win11. ** Please like, comment, and subscribe for more! ** Links to Amazon and Newegg are typically monetized on our channel (affiliate links) and may return a commission of sales to us from the retailer. This is unrelated to the product manufacturer. Any advertisements or sponsorships are disclosed within the video ("this video is brought to you by") and above the fold in the description. We do not ever produce paid content or "sponsored content" (meaning that the content is our idea and is not funded externally aside from whatever ad placement is in the beginning) and we do not ever charge manufacturers for coverage. Follow us in these locations for more gaming and hardware updates: t: http://www.twitter.com/gamersnexus f: http://www.facebook.com/gamersnexus w: http://www.gamersnexus.net/ Our policies, processes, and ethics statements relating to review samples, advertising, travel, errors, and more are transparently and publicly available on this page: https://gamers.nexus/ethics-statements Steve Burke: Host, Test Lead, Writing Patrick Lathan: Testing, Writing Mike Gaglione: Testing Vitalii Makhonvets: Editing [Source](https://www.youtube.com/watch?v=ovOx4_8ajZ8)

Source quotes

[13:21] Steve Burke: 我们目前打心底讨厌 Windows 和微软的间谍流氓软件。因此做这件事是值得的，这项投资完全值得。 [17:27] Gamers Nexus: what we care about is just simply how do the GPUs perform if you're using Linux as your OS so that you can choose between one GPU and the other just like you do with Windows. [21:49] Gamers Nexus: Even the RX 960 XT is a better experience than the 5090, as it's lows are more proportional to the average.

Source quotes

[3:25] Steve Burke: And if more users grit their teeth through some of the initial learning pains, we think that might mean that we can kind of escape from some of the Windows issues as a community and we'd count that as a win.

Original transcript

[0:00] We don't give Microsoft nearly enough [0:02] credit. It's a really impressive [0:03] business model they've built, which is [0:06] spying. It it data acquisition and [0:09] transaction. But look, this has made the [0:12] obstacle to moving to Linux a lot less [0:15] towering of an imposition than it might [0:17] have previously seemed. And it's time to [0:20] start benchmarking Linux, to start [0:22] getting some data out there and [0:24] available for those looking to jump ship [0:26] off of Windows. [0:28] Sorry, I'm not I'm not trying to flex [0:31] right now. Let me just let me just hide [0:35] that off camera. With the unchification [0:37] of Windows and the rise of Steam OS, the [0:39] request for Linux testing are coming [0:41] from more than just a few passionate [0:42] level one text viewers. [0:45] >> Thanks, Steve. [0:46] >> It's been a lot of work, but we finally [0:47] have some Linux benchmarking data for [0:49] GPUs, specifically on the Basite [0:51] distribution. This took about three [0:53] weeks of non-stop testing plus about two [0:55] to three weeks prior to that of bench [0:57] setup research and what we call bench [0:59] hygiene. So researching how you maintain [1:01] a bench through the process of testing. [1:03] Uh we're still learning for all this and [1:05] benchmarking anything to the level of [1:07] isolation we maintain and sort of [1:08] require for our tests. It's just it's a [1:10] lot of work and it takes a lot of [1:12] understanding the controls and [1:13] experience. So you'll kind of have to [1:15] work with us as we go through these [1:17] processes. But maybe this is uh [1:19] insightful for some of you. The first [1:21] round of testing has largely been for [1:23] our own training. That's what we're [1:24] going to present today. We ran into a [1:26] lot of bumps in the road and basically [1:29] plowed through them all. So, we'll [1:30] detail some of those. We're not [1:32] firsttime Linux users, but I mean, for [1:34] me personally, the last time I daily [1:36] drove a Linux operating system was [1:38] probably about 2009 till about maybe [1:41] 2012 or so. And I haven't done a whole [1:43] lot with it except for uh some NAS type [1:45] stuff since then. We spoke with some of [1:47] the basite maintainers during this as [1:49] well and that was helpful for getting [1:51] some of the testing controls dialed in [1:53] and understanding some of the practices [1:55] and best scenarios we might want to [1:57] consider for maintaining a clean [1:59] comparable benchmark environment between [2:01] data sets. So let's get into it. We're [2:03] directly funding our first Linux GPU [2:05] benchmarks with your help from the GN [2:07] store. This content is brought to you by [2:09] you, our audience. Our brand new retro [2:11] themed drink coaster packs launched on [2:13] store.gamersex.net [2:15] and include four distinct coasters per [2:17] pack, all designed by our team in [2:19] Blender, and you can find the link [2:20] below. The coaster packs include an old [2:22] school computer case front panel with a [2:24] floppy disc in the drive slot, a classic [2:26] colorway game controller, an IBM model [2:29] inspired keyboard numpad, and a GN logo [2:31] with all the dials and classic PC case [2:34] ridges and knobs we know from decades [2:35] past, back before AI and first party [2:38] spyware ruined everything. These coaster [2:40] packs are easily cleaned under running [2:42] water and have been fine-tuned to [2:44] balance drinks and cans while making an [2:46] easy condensation catch. Plus, they just [2:48] look cool. We launched these alongside [2:50] our new unhinged vapor wave smiley [2:52] t-shirt, which is brand new and on the [2:54] store now, featuring a dystopian [2:55] censored speech box smiley. You'll see [2:57] this show up in our upcoming [2:59] investigation into a raid of a private [3:01] citizen's home in connection with Sega [3:03] and Nintendo development consoles and [3:04] cartridges, something we flew to the UK [3:06] to investigate. That deep dive special [3:08] report is also supported by our AI [3:10] dystopia report series fundraiser on the [3:12] store, including four high-value packs [3:14] of GN store items and digital rewards, [3:17] some including the coasters. Head over [3:18] to store.gamesac.net to support our [3:21] first Linux GPU benchmarking and our [3:23] investigations all at once. Back to the [3:25] video. Quick dose reality first. Linux [3:27] still isn't for everybody. It's going to [3:29] depend what you do. Some users who [3:31] require certain professional [3:32] applications, for example, may find that [3:34] you just have to use Windows and there's [3:36] no other option because the software you [3:37] use hasn't moved over or it doesn't port [3:40] cleanly or something like that. That's [3:41] the case for our own production machines [3:43] that we use for a lot of the work we do [3:44] here. So, there's systems that we can [3:46] run Linux on like things that might only [3:48] really access web interface software [3:50] that works basically anywhere, but a lot [3:52] of the other stuff it still requires [3:53] Windows. So, that's going to be a [3:55] limiting factor for a lot of people. But [3:56] gaming has gotten a lot more compatible [3:58] on Linux over the last several years [4:00] specifically. And if you're one of those [4:02] web application heavy users with gaming, [4:04] then it might actually be viable for you [4:07] now. Uh and even for professional users [4:09] where there's just more um compatibility [4:12] for your software, that may also be the [4:13] case. So uh it is not difficult in the [4:16] very least to throw a Linux distribution [4:18] onto just a spare SSD or something, [4:20] throw it into a machine and see how it [4:22] goes. And if it doesn't work out, you [4:24] can always go back to your Windows [4:25] system or dual boot or something like [4:26] that. And uh at least it's worth, we [4:29] think, experimenting with Linux at this [4:31] point because I mean, Windows is really [4:33] getting kind of nasty with the level of [4:35] spying that it's doing and how invasive [4:37] it's gotten. Uh and moving away as much [4:40] as it can from bypasses to things like [4:42] always online connectivity, which I [4:44] personally think for an operating system [4:46] is absolutely horrifying and should not [4:49] be tolerated or accepted. uh and this [4:52] may be a good time to start considering [4:53] the jump. So anyway, during this [4:55] process, we had crashes, we had freezes, [4:58] issues with anti-che, some insanely long [5:00] like halfhour long shader compilation [5:02] times. And uh in spite of all of this, [5:05] Linux is still working better than it [5:07] ever has in its history for gaming. And [5:10] if more users grit their teeth through [5:12] some of the initial learning pains, uh, [5:14] we think that might mean that we can [5:17] kind of escape from some of the Windows [5:20] issues as a community and we'd count [5:22] that as a win. Just know what you're [5:23] getting into. It it does work better [5:25] than it used to. It's not perfect. [5:27] Important disclaimer here first is that [5:28] our level of testing for GPUs and CPUs [5:30] on Windows is extremely refined and [5:33] rigid. We've also been highly automated [5:36] for almost a decade now. We don't talk [5:38] about that a lot, but that includes [5:39] capturing images during test passes for [5:41] things like image quality checks and a [5:43] whole lot of cool testing automation and [5:45] data management automation that we just [5:47] we we don't talk about. It's just [5:48] something we do and then we do lots of [5:50] manual vetting of it, make sure it's all [5:52] good. Uh we go through and do reruns [5:54] manually as necessary and we get it into [5:56] the charts. This doesn't exist for us in [5:59] Linux and to refactor it all for Linux [6:01] is not a fast task. So for this we went [6:04] entirely manual which is fine. Uh, and [6:06] we still do that for a lot of stuff. It [6:08] just means two things. One, it's going [6:09] to take longer to do. That's fine. We [6:10] can deal with that. Uh, two, the data is [6:13] not going to be perfectly comparable, [6:14] but it's not anyway because you're [6:15] looking at Linux and Windows. These are [6:17] different environments. Sometimes the [6:18] games themselves are built natively for [6:20] different operating systems, plus [6:22] potential API changes where you might [6:24] need to pivot to a different API to run [6:26] better. There's a lot of problems. You [6:27] can't even cross compare because the [6:29] software that captures the data itself [6:32] is different. This is something a lot of [6:34] people forget about. So the point of all [6:35] this is that we don't want you today to [6:38] go cross compare numbers in these charts [6:40] to numbers in our Windows charts unless [6:43] it's one we've handpicked to do so [6:45] because we know that it is at least [6:47] somewhat comparable. You really [6:49] shouldn't be cross comparing them. [6:51] People are going to do it anyway. I [6:52] know. But you know even the capture [6:55] software measures frame rate [6:58] differently, frame time differently [7:00] between the utilities that exist. Uh and [7:03] so these are things that cannot be [7:04] linearly compared. But the point is [7:06] we've taken a lot of care to do all this [7:08] justice. We just want to be fully [7:09] transparent that Linux is almost by [7:12] design a less rigid solution than [7:15] Windows. It can update more frequently [7:17] or at least it has more components that [7:18] update more frequently and this is [7:20] something that has more of an impact on [7:21] testing. Testing is sort of different [7:23] from a real use and that you need to [7:25] maintain a specific environment uh in [7:27] order to do it properly for comparison. [7:29] So methodology is important here since [7:31] it's all new and sensitive to change. [7:32] The first thing again you need to know [7:34] don't directly compare with our Windows [7:36] testing other than what we've compared [7:37] for you. And secondly with testing we [7:39] have to strike a balance where we want [7:40] to be thorough but we also have to be [7:41] able to complete the test data set in an [7:43] amount of time where there aren't too [7:45] many updates to all the different pieces [7:47] of the software to where it's starting [7:48] to invalidate things. So that's the [7:50] struggle. GPUs are the most [7:51] straightforward hardware category that [7:53] we test. So we began with our standard [7:54] GPU test suite on our standard platform. [7:56] We knew we needed to start with a single [7:58] dro. The front runners were Steam OS, [8:00] which we already have experience with on [8:01] the deck, Cachios, or Basite. Steam OS [8:04] was our first choice, but it's not [8:06] officially available for desktop, so [8:07] that ruled it out for now. Cachios [8:09] should have more cuttingedge performance [8:11] than Basite and might be more flexible, [8:13] but potentially at the cost of some [8:15] stability. Basite is popular. It's built [8:18] for gaming. Thousands of you requested [8:20] we test it specifically, and that's the [8:22] most important part to us. Its [8:23] developers have made executive decisions [8:25] that simplify our own choice making and [8:27] they've even contacted us. And most of [8:29] all, Wendell of Level One Tech even [8:31] recommended it as a stable test platform [8:33] for us to get started with. [8:34] >> The shortest answer I can give you is [8:36] probably Bazite. Basite is good because [8:38] it is an immutable DRO and that is [8:41] important because it is a big [8:42] philosophical difference in how the [8:44] system is managed. A lot of the Linux [8:45] enthusiast community is used to tuning [8:47] and tweaking things and finding ways to [8:49] optimize or minmax stuff or fix problems [8:52] that arise. Uh this we have to address [8:54] right away because testing's a little [8:56] different from use where we need a test [8:59] platform that's predictable and is [9:01] stable and most importantly we have to [9:03] minimize our own judgment calls wherever [9:05] possible because we don't want something [9:08] that is open to change because it was a [9:11] human decision. And as an example of [9:14] that, if you start to minmax and [9:16] optimize or fix bugs for how one device [9:18] runs, now we really need to put that [9:20] equal level of effort into all of them. [9:22] And that's it's just not viable or we [9:25] don't have the experience to do it [9:26] properly. That's the part I'm really [9:28] more worried about. So that's why for us [9:29] testing something closer to usable [9:31] out-of-box state is going to help [9:34] railroad us into a fairer configuration [9:36] for all vendors as we can always fall [9:38] back on the rigid decision-making where [9:40] it exists. Now, Basite Desktop Edition [9:42] comes in at two main versions, Nvidia [9:45] and non- Nvidia, [9:49] just like the S&P 500 at this point. [9:51] After speaking with both the devs and [9:53] Wendell of Level One Tax, we set up one [9:55] drive for each. The OS comes fully [9:57] stocked with drivers and game launchers, [9:59] so we didn't need to do any significant [10:01] configuration after installation, which [10:03] also means that our test benches reflect [10:05] the default user experience as much as [10:07] possible. You still have to be careful [10:09] with crosscontamination, but it's by [10:11] doing different things. Basite is an [10:13] atomic bootable container image. There's [10:15] a full explanation on the FAQ page, as [10:17] well as advantages versus Steam OS. But [10:20] the bottom line is that you can just [10:21] grab an image and slap it on your [10:23] system. Basit's goals are stability and [10:25] compatibility and recoverability in case [10:28] those first two don't work out for you. [10:30] It's intentionally not the highest [10:32] performance most bleeding edge OS, but [10:34] it's a lot lower friction for beginners [10:36] and it's a lot less prone to oh [ __ ] [10:38] moments where I don't know it turns out [10:41] that some feature was broken during [10:42] tested without our knowledge or [10:43] something. According to Antheas from the [10:45] Basite dev team, he said to us, quote, [10:47] we used to include some gaming oriented [10:49] kernel optimizations, but as of August, [10:51] we do not anymore. We found that most of [10:53] the testing for those changes involved [10:55] running, for example, Rocket League at [10:56] 600 Hz while doing something like [10:58] compiling a kernel. Nobody does that or [11:00] cares. Just going to interrupt to say uh [11:03] bold assumption of the Linux community. [11:05] We await your comments post haste. [11:07] Anyway, other than that, he said, quote, [11:09] "We found those optimizations could [11:11] cause a 2 to 30% performance degradation [11:14] where the higher end was on Intel CPUs [11:16] with heterogeneous cores. This actually [11:18] ended up hurting our reputation, which [11:20] is very unfortunate because the version [11:22] of one of the optimizations we used had [11:24] a bug. Moving forward, we will no longer [11:26] make these sorts of changes. So [11:28] performance on Basite should mirror [11:30] Fedora and Arch but with more stability [11:32] as we skip minor versions and delay the [11:35] major a bit. End quote. During a test [11:37] cycle for Windows, we freeze the updates [11:39] for the period of that test cycle. It [11:41] typically takes us about a week. We're [11:42] pretty good at it now. Uh but even if we [11:44] want to run that data for longer, we've [11:47] found through validation, calibration, [11:49] lots of cross comparison, and update AB [11:52] testing over a year even that a lot of [11:54] times there's not big differences. It [11:56] really depends. Sometimes Windows ships [11:58] a major change like for CPUs more than [12:00] GPUs, but for the most part, the data [12:02] can remain fairly shelf stable, so to [12:04] speak. In the very least, we can pause [12:06] the updates for the one week we need to [12:08] test it without major downsides. Now, [12:09] with Linux, there are a lot more layers, [12:12] and those updates may come at a much [12:13] faster pace and sometimes asynchronously [12:16] with each other. Plus, our testing is a [12:18] lot slower. One, because it's not [12:20] automated, and two, because it's just [12:22] slower. So our instinct is to freeze [12:25] versions as much as possible without [12:27] needing to mod things. Linux has a [12:29] number of pieces of software though that [12:31] regularly update. These include the [12:33] operating system itself, it's basite [12:35] here, the games which get potentially [12:37] different updates from Windows, Steam [12:39] which gets meaningful updates as well [12:41] for Linux, the Proton translation layer [12:44] is really important, the drivers among [12:45] other things. Freezing some updates but [12:48] not others. Like if there's something we [12:49] just miss, we don't know it exists and [12:51] we don't freeze it as a result. That [12:53] could lead to issues where things break [12:55] for testing or for a particular device, [12:58] but not for an end user. Our compromise [13:00] has been to freeze basite, including [13:02] drivers and Proton during the test cycle [13:04] only, and take detailed notes on [13:07] everything else, including all version [13:08] numbers for everything we know about. In [13:11] the future, we may shift to GOG, which [13:13] allows using specific game versions. In [13:16] general, we've accepted that we'll have [13:17] to throw away more data and start from [13:19] scratch more often on Linux. We don't [13:21] talk about the business side of running [13:22] the benchmarks in the channel too much [13:24] here. Uh I think it's maybe helpful [13:25] context for this one. So for us, [13:28] typically we would if we didn't do it [13:30] the right way, lose money on a full GPU [13:34] test we passed. Meaning if we took 26 [13:36] GPUs and we ran them through a new suite [13:39] of games with new benchmark passes, new [13:42] data sets, new validation, calibrating [13:45] everything, all of that. If we did all [13:46] that and then we published one video at [13:48] the end of it, which is what we're doing [13:50] here today for Linux, we would lose [13:51] money. It would not be viable. The way [13:53] it can work though for us uh is because [13:56] on the business side, we can then append [13:59] to that data set. So if we run 26 GPUs [14:01] in December and then in January we want [14:03] to review the RTX 5090, the 5080, the [14:06] 5070, whatever, we can append those one [14:09] at a time as they come out. And that's [14:11] where it starts to sort of recoup that [14:13] cost so that it's viable. That's the [14:14] boring business side of it. If you want [14:16] some numbers, for instance, uh we spent [14:18] I went and looked it up 85 hours [14:20] overhauling our GPU testing suite in [14:22] December of 2024. That included new [14:24] games, benchmark validation, [14:26] calibration, running about 26 GPUs [14:28] through it, experimenting with new [14:29] methodologies and getting the plan in [14:31] place for stuff like animation error [14:33] metrics, and then working through to [14:34] eliminate bad data, calculate things [14:36] like standard deviation, all this stuff, [14:38] stuff you don't really see other than [14:39] the charts. If we were to then only [14:41] publish a single video with that 85 [14:43] hours of work plus the editing work, [14:45] plus the camera work, plus whatever my [14:47] scripting time is, that's not going to [14:49] be sustainable. But because of the way [14:52] the data is sort of designed to be shelf [14:54] stable for a little while like at least [14:56] a quarter typically and we do validate [14:58] constantly to make sure it's not [14:59] drifting that makes it work with Linux. [15:03] The problem is that it's not quite as [15:06] easy to hold it in place. So for that [15:08] reason simply rerunning all the tests [15:10] every time we want to add a device for [15:13] Linux benchmarking is not going to work [15:14] for us. What we're probably going to [15:16] need to do for now at least until we [15:18] develop the the processes for this to be [15:20] more shelf stable is going to be running [15:22] Linux test maybe a couple times a year [15:24] at most for GPUs and collecting all new [15:27] data as fast as we can in that period. [15:30] The speed matters because as stuff is [15:32] updating other things can break and we [15:34] might not know if it's because of the [15:36] testing or if it's because it actually [15:38] broke. Anyway, that's sort of some of [15:39] the boring business side of it. Uh, the [15:41] only reason I'm saying this, it's not [15:42] because like pity us, we have to [15:45] benchmark video games and it takes a lot [15:47] of time. It's fine. Like that's what we [15:49] do. We like doing it. Uh, I'm putting it [15:51] out there because I just want everyone [15:53] to be aware that currently our processes [15:56] for this are not as mature and it might [15:57] not be possible to get them to the level [15:59] that they're at for Windows in the near [16:02] future for us. Uh, and because of that, [16:04] it makes it difficult to do a lot of [16:07] these types of tests frequently because [16:09] I mean, it ate three weeks of Mike's [16:11] time running those benchmarks and I'm [16:14] working through all the issues. Now, we [16:15] can speed that up a lot like by choosing [16:16] things that don't have bugs, but you get [16:18] the idea, I think, and uh, we're going [16:20] to keep working on it. It's going to [16:21] take us some time to get it optimized, [16:23] but we hate Windows and Microsoft at the [16:25] moment for the spyware. So, it's worth [16:27] doing. It's worth the investment. All [16:28] right. To cover the equivalent settings [16:31] that we'd change on Windows, we set [16:32] adaptive sync to never, allowed screen [16:35] tearing for full screen windows, and set [16:37] the power profile to performance. We've [16:38] had warnings from Wendell and others [16:40] about colonel spin lock performance [16:42] loss, the CPU performance governer, and [16:44] plenty of other Linuxy type things. But [16:46] the advantage of a plug-and-play OS like [16:48] Basite is that we can just install it [16:50] and run it. Any issues are a valid part [16:52] of the experience at that point. There [16:54] are a lot of ways to [ __ ] this up big [16:56] time. We think we've protected against [16:59] them. We'll find out. I'm pretty sure [17:02] we've done a decent enough job of [17:03] protecting against them in most of this [17:05] testing. Is that enough qualifiers for [17:07] you? Can you tell that we don't have a [17:09] lot of experience publishing Linux [17:10] benchmarks? It's a lot different, even [17:12] though we know how to run benchmarks. [17:13] So, we're starting with our existing GPU [17:15] suite for the most part, including the [17:17] same settings we use for Windows [17:18] wherever possible. They're not always [17:20] all there. We're going to limit the [17:21] amount of Linux and Windows comparisons. [17:23] The frame rate data is collected [17:25] differently. It is Mango HUD we're using [17:27] for Linux versus PresentMmon for Windows [17:30] and uh that feels alone like an apples [17:33] to oranges comparison both because of [17:35] the difference in tools and the [17:36] difference in platforms. I think a lot [17:38] of people want them to be compared [17:39] directly but really what we care about [17:42] is just simply how do the GPUs perform [17:45] if you're using Linux as your OS so that [17:46] you can choose between one GPU and the [17:47] other just like you do with Windows. How [17:49] Linux handles frames specifically is a [17:51] little different. For example, we always [17:53] allow screen tearing, but the concept of [17:55] tearing is still new to Whan. The scope [17:57] of our GPU test suite doesn't include [17:59] extended stability testing. We'll report [18:00] any crashes or incompatibilities that we [18:02] encounter, but we're not going to [18:04] validate that Dragon's Dog 2 can run for [18:06] 5 hours straight on Linux without [18:07] crashing. So, be aware there may be [18:09] issues beyond what we're reporting here. [18:10] Our focus today is on numbers, not [18:12] stability. We're going to go through the [18:14] numbers. We have some more critical [18:16] methodology discussion and issues we've [18:17] run into to get into. I'm going to put [18:20] that after the charts because I know [18:22] people are skipping to the charts. Um, I [18:24] just asked that if you are going to like [18:25] go post this stuff on Reddit or [18:27] something to please watch the parts [18:29] explaining some of the limitations, the [18:32] disclaimers, the issues, things like [18:33] that to have a fuller picture than just [18:35] the numbers. One other thing, we're only [18:36] testing some of the most modern cards [18:38] today. There's a lot of reasons for [18:39] that. The main one is that uh it [18:41] sometimes requires different OS builds [18:42] entirely and we're trying to limit how [18:45] many times we're rebuilding the OS and [18:47] the test platform because it's a huge [18:48] time sync. So, we're going to focus on [18:50] the 50 series and on things that are [18:52] compatible with the builds that we have. [18:53] We'll worry about going back for other [18:55] stuff later. And again, do not cross [18:56] compare to our Windows data because it [18:58] is not necessarily tested the same way. [19:01] It could even be a different API [19:02] entirely. So, that's all important [19:04] stuff. Let's get into the numbers, then [19:05] we'll address the issues during testing. [19:06] For the GPUs on the charts that have [19:08] various VRAM capacities, we used 8 GB [19:10] for the RX960 XT and 16 GB for the RTX [19:14] 560Ti. The only reason for this [19:16] discrepancy is simply due to GPU [19:18] availability as we managed multiple [19:20] other tests during all of this. And [19:22] we'll continue to look at adding more of [19:24] these if we do another round of Linux [19:26] testing. But for now, we'll start with [19:27] these. We'll start with the Windows [19:28] build of Dragons Aqua 2 at 1080p max [19:30] settings. For this testing, we used KD [19:31] Plasma 6.4.4, KD Framework 6.1 7.0, QT [19:35] 6.9.2 Wayland, and 10.0-2E Proton with [19:39] the exception of the 970 XT and the [19:40] 570Ti, which ran earlier in the test [19:42] sequence and were both experimental [19:44] 10.0. 0- 2025 September 19. We've got a [19:48] bunch of other notes as well, but we're [19:50] just we're not going to run them through [19:50] them all today. The 5090 leads the chart [19:53] for 1080p at 176 FPS average. Just like [19:56] with Windows, we'll need to look at [19:57] higher resolutions to ensure it's not [19:58] CPUbound. The 970 XT is disturbingly [20:01] close to Nvidia though, for Nvidia at [20:03] least. Up at 140 FPS average, and [20:06] leading the RTX 5080s 134 FPS result by [20:09] 3.9%. Lows are about the same between [20:12] them. The 5070Ti suffers somewhat in [20:13] 0.1% lows. On inspection, it had one [20:16] pass with a 23 FPS, 0.1% low, one with [20:19] 61, and two at 96 to 97 FPS. These are [20:22] averages, by the way. So, the GPU [20:24] experienced more variance than typical. [20:26] The 9070 is about tied with the 5070Ti, [20:29] but had superior lows. The 9070 also [20:31] leads the 5070 by 20%. The 9060 XT [20:35] outdoes the 5060, but the 5060 Ti 16 GB [20:39] outdo both. 1440p had the 5090 at 154 [20:41] FPS average, leading the 5080 by 37%. [20:44] The 9070 XT runs at about the same level [20:46] of performance as the RTX 5080, which is [20:48] another good position for AMD despite [20:51] seeing a decrease in its relative rank [20:52] from 1080p. The 5070Ti trails this with [20:55] lows proportionally scaling. Current [20:57] Newegg prices have the 5070Ti at $730 to [21:00] $750 or so. That's going to change [21:01] probably by the time this goes up, but [21:03] that's the range it is right now with [21:04] the 970 XT at $600 to $650. In this [21:07] situation, the 970 XT is improved on the [21:10] 570Ti in both performance and price. As [21:12] for the 9070, which is around $550 to [21:15] $570 on Newegging, as we write this, [21:17] it's outdoing the $500 RTX 5070 by 19%. [21:21] The 5060 8 GB falls off an absolute [21:23] [ __ ] cliff here and becomes [21:25] unplayable. 4K draws the numbers down. [21:26] The RTX5090 now has a 44% lead over the [21:29] 5080 as compared to 37% at 1440p, while [21:32] the 5070Ti is roughly tied with the 9070 [21:35] XT. The 9070 is just behind the 970 XT, [21:38] which improves on the nonXT by only [21:40] about 10%. The 5070 ends up below the [21:43] 9070 with the 9060 XT at the bottom of [21:46] the chart. The 5060 was not capable of [21:48] running this test. Blackmth Wukon is [21:50] next. In Windows, this game heavily [21:52] slants in favor of Nvidia with ray [21:53] tracing enabled. It's more balanced in [21:55] rasterization, which is what we'll look [21:57] at first, but Nvidia is still generally [21:58] a strong performer here for Windows. in [22:01] Linux. And at 1080pi with this chart, [22:03] the RTX 5090 runs at 115 FPS average, [22:06] but it has extremely disproportionate [22:08] lows that slam its 1% and.1% averages [22:11] through the floor. These aren't anywhere [22:13] close to the actual average FPS, [22:15] indicating a bad experience despite a [22:17] high average frame rate. We'll look at [22:18] that in a frame interval plot in a [22:20] moment. Unfortunately, animation error [22:22] metrics don't yet work in Linux software [22:24] that we're aware of, or at least not in [22:25] the same way they do in Windows with [22:27] PresentMmon, so we can't evaluate that [22:29] part right now. But we can still look at [22:31] the frametoframe intervals. The 9070 XT [22:33] has a lower average frame rate, but [22:35] offers a significantly better experience [22:37] than Nvidia's flagship RTX5090. The 5090 [22:40] has an average that's 9% ahead of the [22:42] 9070 XT. Not that much. Yet, the frame [22:45] pacing on the 9070 XT breaks away and [22:47] also explains why the average isn't much [22:50] more ahead. The 9070 nonXT repeats this. [22:53] It's only about a 5 FPS difference in [22:54] average from the XT and still retains [22:56] better interval pacing than a 5090. The [22:59] 5080 follows the 5090s behavioral [23:01] patterns, as does the 5070Ti, the 5070, [23:04] the 5060Ti, and the 5060. This is just [23:06] an Nvidia problem with this combination [23:08] of game, driver, and operating system. [23:11] Even the RX 960 XT is a better [23:14] experience than the 5090, as it's lows [23:17] are more proportional to the average. [23:19] For that matter, Intel's B580 is more [23:21] consistent. Nvidia has issues with its [23:24] flagship title with this combination of [23:26] software. And for 1% and.1% lows, it's [23:28] not just about the absolute number that [23:30] they are, but how close that number is [23:32] to the average frame rate. That's what [23:33] we mean by consistency. Here's a plot of [23:36] the frame to frame intervals, previously [23:37] known as frame times. We've been [23:39] fine-tuning some of the language on this [23:41] since our animation error white paper we [23:43] released. The frame frame interval in [23:44] milliseconds for the RTX 5090 [23:46] illustrates frequent spikes to 18 to 20 [23:48] milliseconds, up from a typical result [23:51] of 8 milliseconds. That's a 10 to 11 [23:53] millisecond excursion frame to frame. [23:55] and it's frequent, meaning that you can [23:56] perceive a slight stutter. This is an [23:58] instance where we'd love to have an [24:00] animation error metric in Linux because [24:02] it'd be a perfect way to drill down into [24:04] the issue further. For a 5090, this [24:06] experience is unacceptable. We know it [24:08] doesn't do this in Windows, so this is [24:10] unique to the situation. The RX970 XT [24:12] plots with extremely consistent frame-to [24:14] frame intervals with almost no [24:16] excursions. The few that do exist are no [24:18] more than 3 to 4 milliseconds from frame [24:21] n minus one or the prior frame. So, [24:23] while the 9070 XT has overall lower [24:25] frame rate average or a higher frame [24:26] interval timing, the experience is far [24:28] better when using it at 1440p, the [24:30] RTX5090 establishes more distance and [24:33] average frame rate from the RX970 XT, [24:35] yet still suffers from dismal 1% and.1% [24:38] low averages. The end result is that the [24:40] 9070 XT has remarkably consistent [24:42] frame-to-frame pacing with some of the [24:43] best proportionality we've seen on these [24:45] charts so far. While the RTX 5090 [24:48] Founders Edition flounders its way [24:50] through pacing, the 9070 XT is superior [24:53] in this test. That may change if Nvidia [24:55] is able to fix whatever the problem is. [24:57] However, a fix would likely reduce the [24:58] average frame rate by pacing the frames [25:00] more evenly. 5090 lows here are [25:03] inconsistent, meaning that they vary [25:05] widely runto run. The other Nvidia [25:07] devices suffer a similar fade like the [25:08] 5080. The 9070 ends up better than the [25:11] 5080 by experience and is functionally [25:13] tied in average frame rate. The lows on [25:15] the 5060 are more even with its average [25:18] as we appear to be hitting some kind of [25:19] floor for the driver. For its price, the [25:22] B580 is doing well as compared to the [25:24] 9060 XT and the RTX 5060. Here's the [25:27] frame pacing for the 1440p results. The [25:29] 5090 shows the same behavior as before, [25:32] just with a higher baseline time [25:33] required at closer to 10 to 11 [25:35] milliseconds. The excursions adjust [25:37] accordingly. Now at 19 to 22 [25:38] milliseconds for the spikes or a [25:40] deviation of about 10 to 12 milliseconds [25:42] frame to frame in some cases. The rapid [25:45] succession of these spikes is what makes [25:46] the experience bad for the player. At [25:48] least in terms of how it actually feels [25:50] to play the game. The 970 XT looks about [25:53] the same as before. The average time is [25:55] a little longer, but the fluidity is far [25:57] better. Here's the 4K bar chart. The [25:59] 5090 handles 4K better than anything [26:01] else. So, we finally found a situation [26:02] where the game load crushes the other [26:04] cards enough that even an inefficient [26:06] and poorly running 5090 can still [26:09] outmatch them. The lows come up a little [26:11] from the prior chart, mostly because the [26:13] inconsistency run to run is just so high [26:15] for the 5090s low-end average [26:17] performance, which is just the nature of [26:18] poorly paced frame intervals. In [26:20] general, the RX970 XT now falls [26:22] technically behind the 5080 in average [26:24] FPS with its lows still better than the [26:27] 5080, but worse than the 5090. That [26:29] said, they're more proportional to the [26:31] 970 XT's average, so they may still feel [26:34] more fluid despite being overall lower [26:36] frame rate. Resident Evil 4 is up next. [26:37] We had to prune more outlier test passes [26:39] for this one than other games just due [26:41] to run-to-run consistency issues. At [26:42] 1080p, the 5090 leads at 360 FPS average [26:45] about 27% or the 5080. We need to see [26:48] higher resolutions for scaling away from [26:50] CPU limitations. The 5080 and 9070 XT [26:52] are clearly unconstrained by the CPU [26:54] though, and the 5090 makes that much [26:56] obvious. The 9070 XT is just behind the [26:58] 5080 in all three metrics, giving the [26:59] 5080 a 7.8% advantage in average frame [27:02] rate. Broadly speaking, this isn't [27:04] looking as favorable for Nvidia as an [27:05] Windows environment. AMD seems [27:07] disproportionately favored here as [27:09] compared to typically to be that close [27:11] to a 5080 is not a good thing for [27:13] Nvidia. The 9060 XT is about 48% ahead [27:16] of the B580. As a result, the 9060 XT is [27:19] similar to the 5060 for performance. As [27:21] an aside, the 0.1% low standard [27:23] deviation is very high in this test [27:24] overall. Although the average frame rate [27:26] was generally remarkably consistent at [27:28] 1440p, the 5090 drops to 316 FPS average [27:31] and now leads the 5080 by 51%. That's a [27:34] huge gain over the 27% previously and is [27:36] mostly thanks to imposing more of a GPU [27:38] bind. The 5070Ti and 9070 XT are about [27:41] tied with each other, but the 5070Ti has [27:43] lower 0.1% low marks due to a lower [27:46] performance pass. This is real data [27:48] though and is representative of the [27:49] experience. So, we've included it. The [27:51] B580 and B570 don't perform well again, [27:54] falling below the RTX 560 and landing at [27:56] the bottom of the chart. The cards are [27:58] cheaper than everything else here, [27:59] assuming MSRP, so it's somewhat [28:01] expected. In the very least, though, [28:02] Intel's frame time pacing performance is [28:04] good here with some of the most [28:05] consistent frametoframe intervals on the [28:07] chart. At 4K, the RTX 590 now clears a [28:09] ridiculous 66% lead over the RTX 5080, [28:12] seemingly continuing to gain as load [28:14] increases. It's able to brute force its [28:16] way to the top of other charts. [28:18] Standouts include the RTX 5060, which [28:20] actually did okay for average frame rate [28:21] at 4K, if not for the awful 1% and 0.1% [28:24] load metrics, which call our attention [28:26] to bad frame frame interval pacing. The [28:27] B580 and B570 are at the bottom again, [28:29] but to their credit are both better than [28:32] the Nvidia RTX 5060 result. The middle [28:34] of the chart sees the 5070 and 9070 [28:36] about 10 FPS apart in averages. [28:38] Starfield is up next. No one plays this [28:40] game, so we figured we'd multiply 0 by [28:42] zero and test it in Linux. The 9070 XT [28:44] did phenomenally here by comparison to [28:46] the 5090. Actually, Nvidia is just [28:48] straight up screwed in this game with [28:49] bad performance as a result of the [28:51] drivers of the game interaction. [28:52] Obviously, a 5090 is just better than a [28:54] 9070 XT in basically literally every [28:56] possible metric objectively. However, in [28:59] actual use in this test, its performance [29:01] is worse than an RX970 and an RX970 XT [29:06] alike. The 5090 is just 19% ahead of the [29:08] 5080 here with the 5070 Ti about the [29:11] same as the 5080 and the 5070 not far [29:13] behind. This is a clear indication that [29:15] something is just broken in this chain [29:17] of drivers somewhere and software uh and [29:19] it's falling on Nvidia. The game was [29:21] overall extremely consistent and it's [29:22] average and 1% lows and AMD comes away [29:25] looking good here overall and has thus [29:27] far held more of an advantage in Linux [29:29] benchmarking as compared to its relative [29:31] positioning in Windows. In Starfield at [29:33] 1440p, the same situation unfolds. The [29:35] 9070 XT leads the charts, followed [29:36] closely by the 90709 XT. These two cards [29:39] remain as close as ever to each other. [29:41] The 5090 is next, then the 5080. This [29:43] time the 5090 leads the 5080 by 24%. [29:46] That's not normal, but it's at least [29:47] better. Broadly speaking, Nvidia is [29:49] getting wrecked here. It's not even [29:50] close. The Nvidia GPUs still [29:52] hierarchically line up against each [29:54] other by name, but do not land where [29:55] expected by actual performance [29:57] differences. At 4K, the situation [29:58] improves slightly for the 5090, but not [30:00] in a way that matters much. The 9070 XT [30:02] is still in front now at 60 FPS average, [30:04] and still with lows that scale as [30:06] expected. The 5090 trails basically [30:08] tying the 9070 XT here. The lows [30:10] difference is within variance of the XT. [30:12] Uh the 9070 flanks the 5090 on the other [30:15] side. So overall, the higher resolution [30:16] has created a scenario where it's [30:18] clawing itself ahead. The 5090 that is, [30:20] but still not by an amount that you'd [30:22] come to expect from the hardware it has. [30:23] This looks like how Intel GPUs sometimes [30:26] feel on Windows. The hardware is clearly [30:28] capable of more, but it can't break out [30:30] of some external box. Speaking of Intel, [30:32] it's not present here because it's just [30:34] broken in this title. We'll look at that [30:36] more in a few minutes. Ste is next. [30:38] first at 1080p. The RTX5090 leads [30:41] everything here for average frame rate, [30:42] including about a 43% lead over the [30:44] 5080. The lows for the 5090 and 5080 [30:47] illustrate a limit to how close they can [30:48] get to the average of Nvidia's higherend [30:50] devices, where we're bouncing off of [30:52] some kind of driver or softer ceiling. [30:54] The 9070 XT and 9070 both carry lows [30:56] closer to the average. Once again, with [30:58] average results for the 9070 XT about [31:00] tied with the 570Ti, but better 1% [31:03] and.1% low averages. The B580 appears at [31:06] 74 FPS average, showing that it's [31:08] playable and competitive with at least [31:10] the 9060 XT. Have we said average enough [31:12] in the last two paragraphs? I thought [31:14] so. Let's move on. One thing here, the [31:16] RTX 5070Ti had extremely high run torun [31:19] variance. That appears to be just a part [31:20] of the performance characteristics for [31:22] this configuration. We didn't see this [31:23] with all of their devices. Normally, [31:26] what we would do is because we have [31:27] enough data to support it, we'd strip [31:29] out the bad results, rerun it all, and [31:31] see if they repeat. And if they don't, [31:32] we call it some kind of test deviation [31:34] or an issue of that specific run, like a [31:37] Windows background task. In this case, [31:39] we're including the data because we're [31:40] still new to Linux testing. This appears [31:42] to be a real result, not some test [31:43] outlier. This plot shows that overall [31:45] the frameto frame pacing of four passes [31:48] is generally inconsistent. The spikes [31:50] you're seeing above each cluster of data [31:52] are undesirable, but the worst one [31:53] clearly happened just before frame 3000, [31:55] where we had a spike to about 46 [31:57] milliseconds. This is long enough that [31:59] you definitely notice even though it is [32:00] a single spike. The 1440p results have [32:03] the 5090 50% ahead of the 5080 in [32:05] average frame rate with the 5080 leading [32:07] a nearly tied pairing of 9070 XT and RTX [32:10] 570Ti. The 9070 XT maintains better [32:12] frame interval pacing, although not by [32:14] as much as some previous results. The [32:16] RX970 is likewise consistent. The B570 [32:20] ran at 43 FPS average and is plotted [32:22] here as well. At 4K, the RTX 590 held a [32:24] 119 FPS average and lows that were still [32:27] disproportionately low as compared to [32:28] those of the 5080. To us, this indicates [32:31] again that there's a ceiling the higher [32:32] performance Nvidia card is bouncing off [32:34] of. For pacing, the 970 XT ran at 64 FPS [32:37] average, roughly tying the 5070Ti and [32:39] landing with an error of its 0.1% lows. [32:41] Predictably, the 960 XT, 5060, B580, and [32:45] B570 can't really run this test. It's [32:47] okay though, we still included the data [32:49] just in case the scaling is useful. [32:50] Cyberpunk 2077 Phantom Liberties up now. [32:52] first rasterized immediately. The first [32:54] thing we observed was the higher [32:55] standard deviation this test with 0.1% [32:57] low standard deviation about 10 to 11 [32:59] FPS which is the highest of these charts [33:01] so far. At least three cards were in [33:03] this range for this test. The game is [33:05] just more variable than in our Windows [33:07] environment. The RTX 590 ran at 195 FPS [33:09] average, leading the 970 XT by 32% in [33:12] average frame rate. However, the 9070 XT [33:14] again has superior lows to the Nvidia [33:16] RTX5090 and in a big way. Fortunately [33:19] for Nvidia, the card is still overall [33:20] playable and the large spikes aren't [33:22] easily noticeable. Because its lows are [33:24] high enough overall, it may be the case [33:26] that the 5090 maintains a better [33:28] experience, but only by brute force. The [33:30] 5080 trails the 9070 XT for average [33:32] frame rate and significantly in low [33:34] averages. The 9070 is close in average [33:36] FPS to the 5080 and advantaged in lows. [33:39] Intel's B580 ran at about 60 FPS [33:41] average, which isn't bad for an ARC card [33:43] in this game. that has it behind the [33:45] 5060 which leads by a 20% in average [33:48] frame rate but they are roughly tied in [33:50] 1% and 0.1% lows at 1440p. The 5090 held [33:53] a 148 FPS average with lows at 104 and [33:56] 90. This is a 46 to 50% lead over the [33:58] 9070 XT and RTX 5080 which have about [34:02] the same average frame rate. Lows are [34:03] also numerically better but they are [34:06] less proportionally spaced from the [34:07] average. The Intel B580 drops to 40 FPS [34:10] due to the heavy load of 1440p following [34:12] the RTX 5060. The 5060 struggles with [34:15] the combination of VRAM limitations and [34:17] Nvidia software specific issues in this [34:18] test on this OS. We don't expect the [34:21] 5060 to be able to play this game at 4K [34:23] to be clear, but other than this result, [34:24] the rest is similar to what we've seen [34:25] previously. Nvidia's results are [34:27] improving at the low end, but are still [34:29] disproportionate in the 5090 as compared [34:31] to the other devices. The 9070 XT [34:33] struggled at 4K, landing at 48 FPS [34:35] average and trailing the 5080 in all [34:37] three metrics. The B580 and B570 both [34:40] massively outperform the 5060, as does [34:42] the 960 XT, particularly in the low [34:44] metrics. We're getting into Balders's [34:46] Gate 3 benchmarking now. This game is [34:47] different from many of the others here. [34:49] It has a Linux native build and a [34:51] Windows build, and its test matrix gets [34:53] more complex by the fact that it has [34:55] Vulcan and DX11. The Linux native build [34:58] had some serious issues that we'll talk [35:00] about in a few minutes, but we still [35:01] tested it at 1080p with the Linux native [35:03] build running the Vulcan API. Nvidia [35:05] runs completely sporadically and [35:08] unreliably. We are well aware that the [35:10] 5090 is below the 5070 and that [35:12] obviously is not how these cards should [35:14] behave, nor does it seem like a normal [35:17] or good test result. But after retests [35:19] and evaluation, the only thing [35:21] consistent is inconsistency when it [35:23] comes to Nvidia with the Linux native [35:24] build of Balders's Gate 3 with Vulcan. [35:26] The 9070 XT ran at 238 FPS average here. [35:29] Its lows weren't proportional to that [35:31] average and are overall less consistent [35:33] than we'd like. And then we saw in other [35:35] games, the 960 XT, it lands predictably [35:37] below that, followed by a host of [35:39] unreliable Nvidia results. Now, the [35:41] tests again produced these numbers and [35:43] it is what happened, but it's clearly [35:45] just broken. Broadly speaking, the build [35:47] is buggy due to issues with reliability. [35:49] We're not listing error bars on this [35:51] one. 1440p is also messed up as you'd [35:53] expect. The results for Nvidia are again [35:55] unreliable and all over the place. It's [35:57] clearly just straight up broken. The [35:58] 9070 XT isn't as consistent as it should [36:00] be, though at least runs at an average [36:03] frame rate that's okay. Let's move on to [36:05] a different build. With the Windows [36:06] build and Basite running DX11, things [36:08] look more normal. The RTX 5090, the 580, [36:10] the 5070 are bottlenecked on CPU [36:13] performance with the 5060 Ti also [36:15] bottlenecking, but not as hard. The 8 GB [36:17] RX960 XT ran at 130 FPS average and had [36:20] lows similar to the RTX 560 8 GB card. [36:24] Both significantly ahead of the Intel [36:25] B580. Even still, Intel manages 79 FPS [36:28] for the average and overall okay lows [36:30] when considering it's generally [36:32] disadvantaged to position. The B570 [36:34] isn't far behind and matches the scaling [36:36] we'd expect from prior Windows scaling [36:38] comparisons of the two Intel models. At [36:41] 1440p, things are the same, just scaled [36:43] down. The top three cards are hitting a [36:44] CPU or other bottleneck with 5060 Ti [36:46] occasionally bouncing off of it, but [36:47] still somewhat struggling on its own. At [36:49] 4K, the 5090 hit 170 fps average, which [36:51] is technically ahead of its 1440p [36:53] result. That'd make no sense in a GPU [36:55] bound scenario when going from 1440p to [36:57] 4K. But in this case, the results at [36:58] 1080p, 1440p, and 4K are all [37:01] bottlenecked on nonGPU components for [37:03] the RTX 5090, possibly also including [37:06] software. The end result is that the [37:08] results will be slightly more variable [37:09] due to different constraints. [37:11] Effectively, all three resolutions are [37:13] identical for the 5090 because of [37:14] external limits. We were able to get the [37:16] 9070XT to run in this one. Its result [37:18] was 103 FPS average, whereas we saw 115 [37:21] with the Linux build and Vulcan. [37:23] although not directly comparable. Moving [37:25] on to rateracing benchmarks, starting [37:26] with Cyberpunk 2077 at 1080p. Nvidia has [37:29] always had an advantage in this test on [37:31] Windows and that sustains here. A lot of [37:33] that is just from the actual hardware [37:34] itself. They're just better with their [37:36] RT hardware right now. The 5090 held a [37:38] 135 FPS average, leading the 5080 by [37:41] 46%. The 5070Ti follows with the 9070 XT [37:45] showing up at 72 FPS average and landing [37:47] between the 5070 and 5070 Ti. Intel's [37:49] B580 actually did well here with its 58 [37:52] FPS average planting it just ahead of [37:54] the 5060Ti. That's a compelling argument [37:56] for Intel in this specific case. The RTX [37:59] 5060 fails completely and cannot run [38:01] this test. 1440p is predictably heavy [38:03] for these cards. The 5080 is the last [38:05] card to technically perform above 60 FPS [38:07] average with the 9070 XT approaching the [38:09] 5070Ti but not meeting it. In this [38:11] situation, AMD does not have an [38:13] advantage in low performance and is not [38:15] more consistent in frame pacing than [38:17] Nvidia. Intel, however, is [38:18] disproportionately good for the value of [38:20] its hardware and approaches the RTX 570 [38:23] while maintaining roughly equal or [38:24] slightly better lows. 4K is too heavy [38:26] for this configuration of these cards, [38:28] but we still ran it. The 5090 maintains [38:29] actually good frame interval pacing here [38:31] with its lows relatively close to its [38:33] average, but still it's at 49 FPS [38:35] average. The next closest card is the [38:36] 5080, and that's only at 30 FPS average. [38:39] We'll just do these last two RT charts. [38:41] In Blackmth with RT medium and FSR [38:43] quality, the 5090 ran at 113 FPS [38:45] average, leading the 5080 by just 30% [38:48] and roughly tying in lows. AMDy's first [38:50] appearance isn't until below the 5060, [38:52] and even then, its low performance is [38:54] disproportionately bad. In this [38:56] situation, AMD is getting comparatively [38:58] crushed. Overall, this is something of a [39:01] reversal from prior charts. AMD has [39:03] disproportionately bad lows and overall [39:05] performance. Nvidia, meanwhile, outdoes [39:07] it. The RTX 5060 struggles in average [39:09] frame rate but did okay in its lows. It [39:11] seems to float closer to whatever [39:13] barrier exists in these scenarios. Dine [39:15] Lite 2 is next with ray tracing. Now in [39:17] this test, Nvidia again outperforms AMD [39:19] significantly. The RTX 570 outdoes the [39:22] 9070 XT here with only the 5060Ti and [39:25] 5060 behind the 9070 nonXT. Overall on [39:29] high-end hardware, most the games we [39:30] tested ran without any major issues on [39:33] Basite here. Uh the Linux gaming [39:35] environment is changing actually every [39:36] day in a lot of cases. And so any [39:38] individual bug we experienced could be [39:40] already gone. It could be gone tomorrow. [39:43] And uh Intel cards and cards with less [39:45] VRAM in particular like the RTX 560 [39:48] seemed to be more likely to have [39:50] problems in testing. Before running the [39:52] test, we were having a hard time [39:53] determining whether people consider [39:54] Nvidia's Linux support to be bad in the [39:57] sense that the cards don't work or bad [39:59] in the sense that people don't agree [40:01] with uh Nvidia's approach and think it [40:05] doesn't align with the spirit of Linux. [40:07] So from Anthas again, quote, in the [40:09] context of Nvidia drivers, open refers [40:11] to the kernel module which is now under [40:13] the MIT license, but the drivers are [40:15] still proprietary. Before the open [40:16] module, there was a closed module. And [40:18] the main difference between them is that [40:20] the closed module was made for preGSP or [40:22] GPU system processor cards. Before GSP, [40:25] the Nvidia driver did a lot of heavy [40:26] lifting and had, for example, 10,000 [40:28] lines of monitor Eid fixes. Stuff like [40:31] this is important proprietary [40:32] information for them. So, the driver was [40:34] closed source, which made a lot of Linux [40:36] users upset and was a legal gray area [40:38] with the kernel code. Now with GSP, [40:40] which is a co-processor on the card, [40:42] they can shove all those fixes in the [40:44] firmware of that co-processor, which is [40:46] closed source, and distribute an open [40:48] driver that communicates with that. It's [40:50] just fancy software engineer lawyering." [40:52] End quote. For gaming, the NVIDIA Basite [40:54] build worked as well as the non- Nvidia [40:56] build, at least with the relatively new [40:58] GPUs we tested. Now, I will say for [41:00] Patrick and myself, it was an awesome [41:02] change of pace to be directly contacted [41:04] by the developers of Basite, Mango HUD, [41:07] and pretty much any open-source tool [41:10] that we mentioned publicly in the course [41:12] of working on this. On Windows, the OS [41:14] and the driver are out of our hands, and [41:16] Microsoft employees only talk to us on [41:17] strict conditions of anonymity. So, it's [41:19] a pretty big change. And on Linux, being [41:21] able to directly ask for information and [41:23] request important features and updates, [41:25] that's pretty cool and it's a great [41:27] sign. Now, keeping it real with [41:28] everyone, it's not perfect, and you need [41:30] to set your expectations. It's still [41:32] worth trying, though. But here's some of [41:33] the issues we ran into. We can thank the [41:35] Steam Deck for normalizing how well so [41:36] many games run through Proton. Although [41:38] we had issues in some instances, it's [41:40] still largely true that it's just a lot [41:41] better than it used to be. As Steam Deck [41:43] users are aware, some anti-che software [41:44] flat out refuses to run, which means the [41:47] games can refuse to run. We dropped F124 [41:49] immediately for this reason. This is the [41:51] only game from our suite that was [41:52] affected by this, but since anti-che is [41:54] usually associated with popular [41:55] mainstream games, this remains one of [41:56] the most significant obstacles to Linux [41:58] gaming. Our two RE engine games, [42:00] Resident Evil 4 and Dragon 2, didn't [42:02] show any options for RA tracing. It's a [42:04] known issue, and it may be possible to [42:06] work around, but not without a level of [42:08] tweaking that we're trying to avoid for [42:10] purposes of maintaining a sterile test [42:12] environment. We're skipping RT tests for [42:14] these two games on Linux until Capcom [42:16] figures it out. A few Steam games have [42:18] native Linux branches, which Steam [42:20] always prioritizes for download on [42:22] Linux. Unfortunately, sometimes the [42:23] native Linux versions suck, not [42:25] necessarily for any reason other than [42:27] the developers decided to make them that [42:28] way. In the case of Total War Warhammer [42:30] 3, the Linux build lags significantly [42:32] behind the Windows version. It looks [42:34] like most users opt to run the Windows [42:36] build through Proton instead. So, that's [42:38] what we ended up testing. Balders Skate [42:40] 3 is another game in our suite with a [42:42] native Linux version, although it only [42:43] came out after we had already started [42:45] testing and it promptly broke desktop [42:47] Linux installs because Larry forgot [42:49] about non- Steam Deck Linux users. This [42:51] distributor side messiness will become [42:53] more common as the Steam ecosystem [42:55] expands. We also found that the full [42:57] screen mode in Balders's Gate 3 would [42:58] always run at the native display [43:00] resolution regardless of in-game [43:01] settings, specifically with the Linux [43:03] build. With Cyberpunk 2077, with our RT [43:06] medium settings profile applied, we'd [43:09] consistently get crashes on the second [43:10] loading screen, specifically the second [43:13] one. We could load a save once, but [43:15] trying to reload or exit to the menu [43:17] would crash the game pretty much every [43:19] time. With our nonRT settings profile [43:22] applied, this didn't happen. We also had [43:24] issues across multiple games that may [43:26] have been caused by low VRAM. For [43:28] example, the RTX 5060 was able to [43:30] struggle through Total War Warhammer 3 [43:32] testing at 1080p, but it completely [43:35] failed to get past the benchmark loading [43:37] screen at any higher resolution. Intel's [43:40] cards had the most problems in general, [43:41] although we were told by Basites's [43:43] developers that Intel GPUs would be the [43:45] most affected by updates. So, pausing [43:48] updates may have prevented some fixes, [43:50] but that's the trade-off we have to [43:51] make. They were only paused for a couple [43:52] weeks here. Using the B580, the native [43:55] Linux version of Balders's Gate 3 [43:56] wouldn't launch. Dragons Dogma 2 [43:58] couldn't run long enough to be [43:59] benchmarked and Starfield couldn't load [44:01] saves. And if Starfield can't load [44:03] something, that's really saying [44:05] something. We also saw flickering and [44:07] artifacts at desktop. None of the games [44:09] in our test suite showed RT as an option [44:11] with Intel except Cyberpunk 2077. And [44:14] when we moved down to the B570, it [44:16] couldn't even run Cyberpunk with RT [44:18] without crashing. Now, to be fair, Cyber [44:20] Punk with RT crashed consistently on the [44:22] RX960 XT as well, but Intel has some [44:25] special considerations here. We also had [44:27] multiple issues that may have been [44:29] related to Mango HUD, which is part of [44:31] the logging software. The most obvious [44:33] was that games would frequently crash [44:34] the desktop when Mango HUD finished [44:36] logging. We've been in contact with [44:38] Mango Hud's developer and have some [44:40] ideas, but for now, we're just sort of [44:41] dealing with it. We also had games [44:43] freeze for up to a minute at a time, [44:46] which could be worked around by tabbing [44:48] out and tabbing back in. We [44:49] intentionally hid the Mango HUD overlay [44:51] during all the testing because of an [44:53] earlier issue on the Steam Deck where [44:55] any overlays would cause stuttering and [44:57] other issues. After a certain point, [44:59] Mango Hud stopped working with the [45:00] Heroic Games launcher, which we were [45:02] using to run Cyberpunk 2077. This was [45:05] alarming mainly because it implied that [45:07] something was updating outside of our [45:09] controls despite being so explicitly [45:12] careful and controlled with all the [45:14] software versions. Although it may have [45:16] been due to free desktop's policy of [45:19] marking releases EOL after two years. [45:22] Someone on Reddit had a related problem [45:23] that pointed us to a fix and then they [45:25] deleted the post. So, good luck. Steam's [45:28] first time shader compilation step took [45:31] an extremely long time, especially with [45:33] Nvidia cards. Up to 25 minutes for [45:35] Starfield, for example, on one of the [45:36] Nvidia cards we tested. Many of the [45:38] Steam games we tested forced tiny shader [45:41] cache updates every day, which is [45:43] behavior we've seen on the Steam Deck as [45:44] well. So overall, we've got some initial [45:46] numbers. Now, for us, this was mostly [45:48] exploratory today. And the biggest thing [45:50] we've learned is that this is going to [45:51] be difficult to get to the level we're [45:53] at on Windows where we can just append a [45:54] card in say a month or two months later [45:57] after collecting the core data set. Um, [46:00] because it changes a lot more. Now, [46:02] there's another compromise we haven't [46:03] talked about here, which is because I [46:04] don't really like it. But another option [46:07] would be to allow the updates to keep [46:08] going, catalog them in some way that's [46:11] transparent, like whatever next to the [46:13] card or something. So if proton changes [46:14] or something uh and then run it anyway [46:17] but now you're cross comparing new data [46:19] to old data and if the percent [46:21] difference swings as a result of that [46:23] but you haven't retested the old stuff [46:25] then it's not accounted for and it makes [46:26] the new test look potentially unfairly [46:29] good or bad performing depending on how [46:31] the update went. So anyway there's stuff [46:32] we can do to further improve this. An [46:34] example would be uh handpicking games [46:36] that specifically run well on Linux [46:39] regardless of any other usefulness like [46:41] comparing to other data stuff like that. [46:43] That might help speed things up a lot. [46:45] Um there's other stuff we could do too. [46:46] We can maybe port some automation [46:48] whatever. So we have ideas we could do [46:49] but in the very least we got it started. [46:51] This has been a huge community request [46:53] for a very long time. I wanted to give [46:55] some context for like why this isn't [46:57] easy to just roll out and why it's not [46:59] quick for us. you could do it quicker, [47:01] but you know, we we really have rigid [47:03] processes to try our best to maintain [47:05] data integrity and for something we're [47:07] less familiar with. That takes a lot [47:09] longer. So, um hopefully it's [47:11] interesting to you and uh you know, [47:13] depending on how the performance does [47:14] and everything, we'll look at um doing [47:16] these more frequently. We can't do them [47:18] as frequently as Windows benchmarks or [47:21] GPUs yet, but who knows? It may get [47:23] there. Thanks for watching. If you want [47:24] to support this stuff directly because [47:25] we could really use your help on [47:26] something like this. This is like it's a [47:28] lot of lot of effort. I'm not going to [47:30] be able to just add a GPU to the chart [47:32] like I was really hoping I'd be able to [47:33] do. It's just it's it's going to be a [47:35] mis, you know, a disservice if we do it [47:38] that way without it all being really [47:40] packed in. So, if you want to help us, [47:41] go to store.k.net and you can grab [47:44] something like our new retroio themed [47:47] coaster packs or our backer tiers for [47:49] our AI dystopia series of special [47:53] reports that we're working on right now. [47:55] And uh that'll help us out a lot while [47:57] you get something cool in return. For [47:59] example, our new unhinged vapor wave [48:01] smiley t-shirts with the censored speech [48:04] box or our YouTube copyright strike [48:06] takedown t-shirt which we dedicate to [48:09] our good friends at Bloomberg. Thanks [48:10] for watching. Subscribe for more. We'll [48:12] see you all next time.