Intel Pitches Xeon Over GPUs to Cut Video AI Costs Sixfold

@Scale

Executing agentic media pipelines directly on Xeon processors cuts token costs more than sixfold over cloud GPUs by eliminating PCIe bus bottlenecks across serial tool calls and state management.

AI-assisted summary of Scaling Media AI: From Pixel-Level Enhancement to Actionable Intelligence on Intel® Xeon® by Intel — original source. Check the original for full context.

Source published: 2026-11-02

Key ideas

Intel showcased high-throughput Video Super-Resolution (VSR) running on Intel Xeon CPUs with Intel AMX, hitting over 900 fps for 1080p-to-4K upscaling and slashing power consumption by 80% compared to equivalent GPU deployments. (3:06)
By inserting an AI-based perceptual pre-processing model optimized with OpenVINO and AMX before SVT-AV1 encoding, Intel achieved a 24% to 28% bitrate reduction while utilizing only four additional CPU cores. (8:08)
The new Intel Media AI Framework integrates Model Context Protocol (MCP) tools, an agentic toolkit, and a visual pipeline designer to unify raw pixel enhancement with semantic understanding and dynamic ad insertion. (12:31)
Adopting a CPU-first architecture for agentic media intelligence eliminates PCIe latency bottlenecks, leverages large system RAM for expanding agent state history, and delivers over six times lower token costs compared to typical cloud GPU instances. (21:06)

Detailed notes

## Background Published on YouTube by @Scale on November 2, 2026, this technical session features Intel's Arnaud Perrier and Richard Chuang addressing the severe infrastructure bottlenecks and runaway costs caused by over-reliance on GPUs for media AI. The presentation outlines how organizations can rebalance video pipelines onto latest-generation Intel® Xeon® processors with Intel® Advanced Matrix Extensions (Intel® AMX), pairing pixel-level workloads with the Intel® Media AI Framework to deliver agentic orchestration, reusable Model Context Protocol (MCP) tools, and video RAG.

## TL;DR The session examines the operational limits of GPU-only video pipelines and presents a CPU-first architectural model for scaling media enhancement and AI reasoning. Intel demonstrates how offloading tasks like Video Super-Resolution (VSR), SVT-AV1 perceptual bitrate optimization, and Whisper audio transcription to Intel Xeon CPUs dramatically slashes power consumption, lowers cloud token costs, and frees up GPU capacity. To bridge raw pixel processing with higher-level semantic understanding, the presenters detail the Intel Media AI Framework, an open, modular ecosystem featuring visual pipeline composition, dynamic inference routing, and optimized MCP tooling. Concrete real-world implementations validate this approach, including dynamic contextual ad insertion and a live multi-modal basketball analysis pipeline running well above real-time broadcast speed on a single Xeon server.

## Atmosphere & style The presentation carries the polished, authoritative cadence of an enterprise systems-architecture briefing, blending measured engineering pragmatism with high-level developer evangelism. Arnaud Perrier delivers crisp, benchmark-driven technical arguments challenging conventional GPU-centric wisdom, while Richard Chuang provides a steady, methodical walkthrough of software design abstractions and developer workflows. Watching the session feels like attending a focused technical summit where abstract agentic concepts are swiftly grounded in concrete silicon telemetry and live video ingestion.

## Bulletin - [03:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=186s) Intel showcased high-throughput Video Super-Resolution (VSR) running on Intel Xeon CPUs with Intel AMX, hitting over 900 fps for 1080p-to-4K upscaling and slashing power consumption by 80% compared to equivalent GPU deployments. - [08:08](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=488s) By inserting an AI-based perceptual pre-processing model optimized with OpenVINO and AMX before SVT-AV1 encoding, Intel achieved a 24% to 28% bitrate reduction while utilizing only four additional CPU cores. - [12:31](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=751s) The new Intel Media AI Framework integrates Model Context Protocol (MCP) tools, an agentic toolkit, and a visual pipeline designer to unify raw pixel enhancement with semantic understanding and dynamic ad insertion. - [21:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1266s) Adopting a CPU-first architecture for agentic media intelligence eliminates PCIe latency bottlenecks, leverages large system RAM for expanding agent state history, and delivers over six times lower token costs compared to typical cloud GPU instances.

## Key takeaways - [00:05](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=5s) Surging video stream volumes driven by generative AI require moving beyond brute-force encoding toward unified pipelines combining computational efficiency with real-time semantic reasoning. - [00:05](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=5s) Intel's media evolution spans hardware-accelerated Quick Sync Video, open-source SVT-AV1 compression, and open libraries like IVSR and Media Transport Library (MTL). - [03:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=186s) Offloading FlashVSR super-resolution to Intel Xeon CPUs enables generative video models (such as Wan 2.1/2.2) to render at 480p on GPUs before CPU upscaling, speeding up end-to-end workflows roughly 7x while saving GPU VRAM. - [03:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=186s) Small Pixels VSR models running on Intel Xeon processors with AMX surpass 900 fps for 1080p-to-4K upscaling at an 80% reduction in power consumption compared to GPU alternatives. - [08:08](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=488s) Traditional SVT-AV1 encoder scaling incurs a 10x CPU complexity penalty between presets M5 and M1 for an 18% bitrate gain, whereas Intel's AI perceptual pre-processing delivers a 24% bitrate reduction using just 4 extra cores. - [08:08](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=488s) A dual-socket Intel Xeon 6 AP server running Whisper Large v3 achieves over 2,000 tokens per second, allowing over 600 hours of speech to be transcribed within a single clock hour. - [12:31](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=751s) The Intel Media AI Framework addresses legacy pipeline brittleness by providing modular, reusable MCP tools and pre-optimized models directly from the Intel Model Zoo. - [13:29](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=809s) The framework's four-tier architecture spans a visual pipeline designer, an MCP tools layer, the Intel Agentic Toolkit, and the underlying Intel Xeon hardware foundation. - [15:43](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=943s) A low-code visual pipeline designer enables developers to prototype multimodal workflows, bind models, and swap MCP components quickly across explore, develop, and scale lifecycle phases. - [16:45](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1005s) The Intel Agentic Toolkit provides operational tooling including prompt complexity-based dynamic inference routing, runtime guardrails, cost-per-million-tokens monitoring, and multi-node scaling via Kubernetes and Ray. - [18:09](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1089s) Dynamic contextual ad insertion combines automated product website asset scraping with semantic keyframe search to insert non-disruptive L-shaped graphic banners into movie streams without interrupting viewers. - [21:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1266s) Agentic AI loops rely heavily on serial logic, JSON parsing, and external API calls—tasks ideally handled by high-speed CPU cores rather than GPU matrix engines. - [21:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1266s) Executing agentic orchestration and inference natively on Intel Xeon eliminates PCIe transfer overhead and leverages host RAM capacity to manage extensive conversational and multimodal agent states. - [22:23](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1343s) A live 1080p30 basketball broadcast demo on a single Intel Xeon 6 processor concurrently performed real-time frame-by-frame shot detection and live commentary transcription while consuming only a low single-digit percentage of CPU capacity.

## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=5s) The Explosion of Video Streams and the Media Infrastructure Challenge

The media industry is undergoing an unprecedented surge in video content, especially live media streams. This volume explosion is heavily fueled by generative AI models and their expanding footprint across modern content platforms. However, delivering this massive scale of high-quality streaming media exposes severe operational bottlenecks. Arnaud Perrier emphasizes that tackling this reality requires moving far beyond brute-force pipelines toward a unified architecture combining deep computational efficiency with real-time reasoning: “Successfully delivering these streams at scale is increasingly challenging and requires a much more efficient AI infused processing pipeline going forward. But even more importantly, a deep media understanding with reasoning in real time so that platforms can further enhance the user experience but also further personalize and monetize that experience.”

Intel’s historical involvement in video processing demonstrates an ongoing progression against compute bottlenecks. The journey began with dedicated hardware acceleration through Intel Quick Sync Video on the CPU to scale encode performance, followed by open-source compression milestones like SVT-AV1, developed alongside the Alliance for Open Media (AOMedia) and recognized with a Technology & Engineering Emmy Award. Yet, as developers started adopting standalone neural networks and AI models for media enhancement, processing pipelines fractured into isolated operational silos. To bridge these disparate islands, Intel introduced open-source projects including the IVSR (Intel Video Super Resolution) library for enhancement workloads and the Media Transport Library (MTL) for raw video transmission over IP in broadcast environments. Today’s infrastructure challenge demands merging AI-enhanced pixel processing with actionable, real-time intelligence directly on mainstream compute platforms.

```mermaid flowchart LR accTitle: Evolution of Intel Media AI Processing accDescr: Progression from hardware encoding to software codecs, isolated AI libraries, and unified AI-infused pipelines. subgraph S1["Legacy Encoding"] A["Hardware Encode (Intel Quick Sync Video)"] --> B["Open Source Software (SVT-AV1 Emmy Winner)"] end subgraph S2["Siloed AI Acceleration"] B --> C["Media Enhancement (IVSR Library)"] C --> D["IP Video Transport (Media Transport Library)"] end subgraph S3["Actionable Media AI"] D --> E["AI-Infused Pipeline (AMX Inference on CPU)"] E --> F["Agentic Media Framework (Real-Time Reasoning)"] end ```

[03:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=186s) Modernizing the Video Pipeline: VSR, Generative Assist, and Open Frameworks

Standard video transcoding pipelines are undergoing a fundamental transformation into AI-infused enhancement pipelines that actively elevate content quality rather than simply compressing raw bits. At the pre-processing and enhancement stages, AI enables Video Super-Resolution (VSR), smoother motion estimation via frame rate conversion, and Inverse Tone Mapping (ITM) to expand dynamic range while preserving subtle visual details. Audio processing has similarly evolved, using automatic speech recognition (ASR) and text-to-speech (TTS) to enrich live streams with automated metadata and live captions. These workflows are measured by modern automated AI quality metrics such as VMAF and UVQ (Unified Video Quality). To eliminate developer friction, Intel integrates these acceleration components natively into major industry runtimes—including PyTorch, SGLang, vLLM, and OpenVINO—and surfaces them through turnkey FFmpeg AI plugins.

Super-resolution forms a critical pillar across broadcasting, streaming, user-generated content (UGC), and FAST (Free Ad-Supported Streaming TV) channels seeking to deliver premium user experiences. Because no single neural model fits every pipeline constraint, developers must balance image quality against computational latency. Collaborating with Small Pixels, Intel demonstrated the performance of optimized VSR models running on Intel Xeon CPUs powered by Intel Advanced Matrix Extensions (Intel AMX). The CPU-based pipeline achieved over 900 frames per second for 1080p-to-4K upscaling, matching native 4K quality while slashing energy consumption by 80% compared to equivalent GPU deployments.

```mermaid flowchart TD accTitle: Generative Video Generation Acceleration via FlashVSR accDescr: Pipeline showing low-resolution generation on GPU handed off to CPU for FlashVSR upscaling to 1080p. subgraph GPU["High-Power GPU (RTX 6000)"] P["User Text Prompt"] --> G["Fast GenAI Generation (Wan 2.1/2.2 @ 480p)"] G -.->|"Saves GPU VRAM & Time (~5 min vs 36 min)"| B["Parallel Clip Generation"] end subgraph CPU["Intel Xeon CPU (with AMX)"] G --> F["FlashVSR Super-Resolution (Upscale 480p to 1080p)"] F --> OUT["Near-Native 1080p Output (7x Faster Total Workflow)"] end ```

The combination of CPUs and GPUs proves equally critical in generative AI video workflows. Generating a 5-second video clip using advanced models like Wan 2.1/2.2 natively at 1080p resolution on a high-end workstation card like an NVIDIA RTX 6000 requires roughly 36 minutes, making large-scale deployment prohibitive. By generating content at 480p instead, the GPU synthesis time drops drastically to around 5 minutes, though the resulting output lacks sharpness. Arnaud explains how offloading high-quality upscaling fixes this tradeoff: “And running FlashVSR on the CPU frees up the GPU memory so that you can generate either longer clips or multiple clips at the same time in parallel on the GPU. So the GPU becomes an accelerator for the entire workflow achieving, you know, in the order of seven times faster generation with the quality that you want for your users.”

- 1080p-to-4K super-resolution throughput: Over 900 fps on Intel Xeon processors using Small Pixels models. - Energy reduction: 80% power footprint savings on CPU compared to standard GPU-based upscaling setups. - Generative video compute: Generating a 5-second 1080p clip directly takes 36 minutes on an RTX 6000; generating at 480p takes 5 minutes. - End-to-end acceleration: High-quality FlashVSR on CPU delivers a ~7x workflow speedup while freeing up GPU VRAM for multi-clip concurrency.

[08:08](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=488s) Perceptual Bitrate Optimization and Audio Inference on Intel Xeon

High-quality software video encoding perpetually struggles with the fundamental tradeoff between compression efficiency and compute complexity. Achieving higher compression to reduce bandwidth delivery costs for premium viral content requires steep, non-linear increases in compute resources. In production deployments of SVT-AV1, transitioning from a balanced operating preset like M5 down to the maximum compression preset M1 yields an 18% bitrate reduction but incurs a 10x penalty in encoding latency and CPU load. To break this scaling barrier, Intel developed an AI-driven perceptual pre-processing model optimized with OpenVINO and Intel AMX instructions. Applied right before the AV1 encoding loop, the model prunes visually imperceptible details.

Arnaud points out that augmenting the encoder with an AI-based pre-processing step accomplishes compression gains that traditional codec algorithms cannot reach alone: “When applied before the encoding stage you can achieve a 24% bitrate reduction over M1 at the same complexity which would be impossible using AV1 encoding technology or just a codec at any preset for that matter and that's only using an additional four cores on the CPU.” When integrated deeper into the SVT-AV1 loop as a QP-aware multi-modal model, the pre-processor delivers up to a 28% bitrate reduction compared to standalone M1 encoding and a 25% reduction compared to M5. Partner ISV Bluedot operationalized this technique in their real-time Perceptual Quality Optimizer on Xeon, reliably achieving 20% bitrate savings on 4K streams without perceptible degradation.

Beyond visual pixels, high-throughput audio inference runs efficiently on CPU infrastructure to power live transcription, localization, and automated monetization. Using open-source models such as OpenAI Whisper for speech-to-text and Opus-MT for multi-language translation, standard Xeon platforms handle real-time multi-stream workloads without discrete accelerators. A single dual-socket Dell PowerEdge server equipped with Intel Xeon 6 AP processors running Whisper Large v3 achieves over 2,000 tokens per second, transcribing more than 600 hours of audio within a single clock hour at reference-grade accuracy. Video technology partner Broadpeak has integrated this Xeon-accelerated stack directly into their software application to deliver real-time, dynamic multi-language subtitle translation.

- Standalone AV1 scaling: Moving from SVT-AV1 preset M5 to M1 requires ~10x more CPU resources to capture an 18% bitrate reduction. - AI pre-processing gains: Intel’s perceptual model achieves 24% lower bitrate than M1 at equivalent complexity, consuming only 4 extra CPU cores. - QP-aware SVT-AV1 integration: Reduces bitrates up to 28% below M1 and 25% below M5 presets. - Bluedot Perceptual Quality Optimizer: Sustains an average 20% bitrate reduction on real-time 4K broadcast video with zero perceived visual loss. - Whisper Large v3 performance: Surpasses 2,000 tokens per second on a dual-socket Xeon 6 AP server, translating to over 600 hours of audio transcribed per hour.

[12:31](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=751s) Bridging Processing and Reasoning: The Intel Media AI Framework

Connecting raw pixel processing with higher-level semantic understanding requires architectural evolution. Handing the presentation over to Richard Chuang, the focus shifts toward unifying compute pipelines with agentic AI orchestration. Historical media workflows have relied on rigid, single-purpose pipelines hardcoded to perform specific tasks such as resizing, closed-caption extraction, or basic transcoding. These legacy setups fail when modern applications require deep context, automated searchability, and dynamic ad insertion driven by contextual understanding.

To eliminate these constraints, Intel created the Media AI Framework, an open, modular ecosystem designed to bridge pixel processing and semantic AI reasoning. Richard outlines the fundamental architectural challenges the framework is designed to eliminate: “Traditionally media AI workload they are hardcoded for specific purpose. With this framework we will provide a number of components that are reusable verified so that we will accelerate the development process of your application. This slides highlights three gaps that we trying to resolve using media AI framework. The first one is to make all the tool and models reusable that I just mentioned. And the second is the model has to be optimized and that takes a lot of work you know for all of the developers and Intel we will provide those optimized model in our model zoo so that our developer can just use it immediately without having to do their own optimization.”

- Gap 1 — Reusability: Replacing brittle, hardcoded media pipelines with standardized, modular components and Model Context Protocol (MCP) compatible tools. - Gap 2 — Optimization: Alleviating the burden of complex CPU inference tuning by providing pre-optimized, verified models out-of-the-box via the Intel Model Zoo.

[13:29](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=809s) Intel Media AI Framework Architecture and MCP Integration

Richard Chuang breaks down the architectural blueprint of the Intel Media AI Framework, detailing how it decouples complex video intelligence tasks into modular, reusable components. Rather than forcing developers to build specialized tools from the ground up, the framework introduces pre-built Model Context Protocol (MCP) integrations that are specifically tuned and accelerated for Intel hardware platforms. These tools enable agentic systems to interact directly with media pipelines, retrieval databases, and model endpoints without custom bridging logic.

The architecture organizes capabilities across four clear horizontal tiers. At the summit sits the control plane and visual pipeline designer, providing access control, registry management, and graphical canvas interfaces for composing agentic workflows. Immediately beneath is the MCP tools and API gateway layer, exposing modular media functions. The third tier comprises the Intel Agentic Toolkit, packing optimized runtime libraries, execution engines, model repositories, and inference backends. Grounding the entire stack is the hardware foundation of Intel Xeon processors. Developers are afforded flexible entry points: they can build client-facing applications against high-level graphical interfaces and RESTful APIs, hook MCP-capable agents straight into the second-tier toolsets, or interface directly with the low-level Agentic Toolkit for full-stack engineering control.

```mermaid flowchart TD accTitle: Intel Media AI Framework Layered Architecture accDescr: A four-tier architectural diagram illustrating control plane, MCP toolsets, Intel Agentic Toolkit, and the underlying Intel Xeon processor platform.

subgraph UserAccess ["Developer Entry Points"] UI["GUI & API Gateway Calls"] MCPApp["MCP-Capable Applications"] ToolkitDev["Full-Control Toolkit Developers"] end

subgraph FrameworkLayers ["Intel Media AI Framework"] L1["Control Plane & Pipeline Designer (Workflow Editing, Registry, Access Control)"] L2["Media AI MCP Tools & API Gateways (Optimized MCP Toolsets)"] L3["Intel Agentic Toolkit (Runtimes, Model Zoo, Optimized Libraries)"] L4["Hardware Foundation (Intel Xeon Scalable Processors with AMX)"] end

UI --> L1 L1 --> L2 MCPApp --> L2 L2 --> L3 ToolkitDev --> L3 L3 --> L4 ```

- Richard Chuang highlights the convenience of pre-packaged ecosystem tools: “We are going to provide a number of MCP tools that is available for you to integrate without having to create you know all different type of MCP tool by yourself and the best part is this MCP tool is also optimized for Intel platform.” - Top layer: Control plane and pipeline designer overseeing workflow composition, security, and registry catalogs. - Second layer: Intel-optimized Media AI MCP tools and API gateways accessible directly by autonomous agent frameworks. - Third layer: Intel Agentic Toolkit supplying runtime engines, curated model libraries, and developer repositories. - Foundation: Native, accelerated execution directly on Intel Xeon server hardware.

---

[15:43](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=943s) Visual Pipeline Designer and Developer Journey

Turning to ease of development, Richard Chuang walks through the visual design environment built into the framework. Prototyping multi-modal video pipelines often bogs down in low-level plumbing code, data serialization, and manual model wiring. The pipeline designer replaces this friction with an interactive visual environment where engineers can drag, drop, and connect verified processing blocks with minimal custom boilerplate.

Developers whose applications need to link directly into MCP tools or tap into pre-validated models in the framework's model zoo can bind components visually. A developer can link keyframe extractors, speech models, and vision-language modules together, wire them directly into registered MCP endpoints, and instantly validate the pipeline flow. The system is designed to accelerate rapid experimentation: if early test outputs reveal quality gaps, engineers can immediately halt the execution, swap out an MCP tool or substitute a different model from the registry, and re-run the loop in seconds.

- Richard Chuang emphasizes the low-code design philosophy: “Pipeline designer let you compose your workflow visually and the prototype using verified components with less custom code.” - Model zoo integration lets engineers bind multimodal models (vision, speech, language, keyframe reasoning) directly to pipeline nodes. - Tight prototyping loops allow engineers to swap out MCP tools or adjust inference endpoints on the fly without rewriting ingestion logic. - The platform covers the entire developer lifecycle across three clear phases: explore, develop, and scale.

---

[16:45](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1005s) Agentic Infrastructure, Inference Routing, and Operational Tooling

Autonomous media agents require substantial supporting infrastructure beyond isolated reasoning models, MCP servers, and vector databases. Richard Chuang details the operational capabilities embedded within the Intel Agentic Toolkit to keep pipelines reliable and cost-effective in enterprise production. A primary feature is dynamic inference routing, which evaluates incoming prompt complexity and semantic context to dispatch requests intelligently—sending lighter queries to compact language models and reserving resource-intensive large models only for intricate reasoning.

Operational management is tightly woven into the runtime environment. The toolkit incorporates fine-grained access control, runtime guardrails, and observability instrumentation. Chuang underscores that operators must maintain constant visibility over metrics like cost per million tokens, ensuring agents remain financially viable at scale. For deployment, the stack provides elastic flexibility: engineers can spin up a quick single-node test environment using Docker Compose, or deploy enterprise-grade, multi-node agent meshes managed via Kubernetes and Ray distributed compute clusters.

- Richard Chuang cautions that modern agentic deployments demand end-to-end operational rigor: “Agents need more than reasoning models, NCP servers and databases.” - Dynamic inference routing dispatches requests across small or large language models depending on prompt complexity to reduce latency and token spend. - Built-in operational tooling monitors guardrails, authorization, and cost-per-million-token metrics across all inference endpoints. - Deployment topologies scale seamlessly from standalone Docker Compose instances up to distributed Kubernetes and Ray multi-node clusters.

---

[18:09](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1089s) Real-World Use Case: Dynamic Contextual Ad Insertion

To demonstrate the concrete business value of deep media understanding, Richard Chuang presents an automated, contextual video advertisement insertion pipeline. The demonstration centers on dynamically placing a commercial advertisement for a juice beverage into a feature-length movie without disrupting the viewer experience. Rather than relying on intrusive, pre-scheduled mid-roll commercial breaks, the framework orchestrates a seamless native insertion.

The Intel Media AI Framework first analyzes the external product by scraping and understanding the juice brand’s public website, generating an automated L-shaped graphic overlay banner. Concurrently, the movie content is processed using the framework's keyframe vision extractors and audio reasoning tools. Running video-to-video semantic search, the framework scans the timeline to identify the exact narrative moment where a juice advertisement fits naturally into the scene context. Once the optimal insertion window is pinpointed, the resulting timestamp metadata is dispatched to an ad insertion engine to project the L-shaped banner dynamically across either linear broadcast or non-linear streaming delivery.

- Richard Chuang highlights the collaborative ecosystem model behind dynamic placement: “The Intel media AI framework provides the media understanding while propic handle the ad insertion supporting both linear and nonlinear placement.” - The automated ad workflow ingests product assets from public websites and automatically synthesizes contextual L-shaped graphic banners. - Keyframe and audio reasoning models locate semantically relevant, non-disruptive placement windows in the video stream. - The targeted placement creates a three-way value proposition, delivering a win-win-win scenario that benefits advertisers, content owners, and viewers alike.

---

[21:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1266s) The CPU-First Imperative for Agentic AI Workloads

Arnaud Perrier takes over the presentation to synthesize the fundamental systems architecture argument: why the central processing unit is uniquely suited as the core compute engine for agentic media intelligence. Perrier clarifies that agentic AI is fundamentally distinct from bulk batch inference or pure matrix-multiplication text completion; it consists of an intensive, continuous loop of serial logic branching, API gateway calls, state machine updates, and external tool execution.

Offloading these composite workflows back and forth between host CPUs and dedicated GPU accelerators introduces severe latency penalties and memory transfer overhead. Running both the control plane orchestration and the inference tasks natively on Intel Xeon establishes unified compute, completely eliminating PCIe bus data transfer bottlenecks. Furthermore, CPUs feature high-speed single-threaded execution cores that excel at parsing complex JSON payloads, running Python orchestration code, and managing tools. In addition, massive server-grade system RAM provides the vast memory capacity necessary to maintain rapidly swelling agent state histories that would easily exhaust accelerator VRAM.

- Arnaud Perrier delivers the core systems architectural verdict: “As a summary, Aentic AI is a continuous loop of serial logic, API calls, state management, and not just simple text generation.” - Unified compute on Intel Xeon eliminates the latency and PCIe bandwidth penalties inherent in shuttling media frames and token buffers to discrete accelerator cards. - High-speed single-threaded execution accelerates control-flow tasks like JSON deserialization, Python script execution, and tool coordination. - Large system memory pools accommodate rapidly growing agent conversation logs and multi-modal state vectors directly in RAM. - Arnaud Perrier outlines the financial upside of this architecture: “The bottom line is to start CPU first to simplify your stack on a unified footprint which offers you a you know in excess of six times lower token costs compared with you know typical cloud GPU instances.”

---

[22:23](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1343s) Live Demo: Real-Time Basketball Visual Understanding and Next Steps

To ground the architectural discussion in live silicon performance, Arnaud Perrier showcases a live visual understanding demonstration running on a single Intel Xeon 6 server. The workload processes a live, uncompressed camera feed from a high school basketball game, driving two parallel pipelines simultaneously across the video and audio streams on the exact same CPU socket without GPU assistance.

The video pipeline inspects every individual frame of the 1080p stream running at 30 frames per second, running computer vision models to detect whenever a player attempts a shot at the basket and instantly generating a timestamped highlight clip card. Simultaneously, the audio pipeline ingests the live announcer commentary, running real-time speech recognition to produce a synchronized running transcript alongside the video feed. A system telemetry dashboard displayed in the upper-right corner indicates that the entire dual-pipeline setup executes several times faster than real-time broadcast speed, consuming only a modest few percent of the total server CPU capacity during real-time 1080p30 ingestion.

- High school basketball game feed is analyzed live at 1080p resolution at 30 frames per second on a single Intel Xeon 6 server. - Dual parallel pipelines execute concurrently on host CPU cores: one tracking frame-by-frame shot detection for highlight reels, the second transcribing commentary audio in real time. - The end-to-end multi-modal pipeline runs multiple times faster than real-time playback, utilizing only a minor single-digit percentage of the Xeon 6 compute capacity during 1x real-time streaming. - Arnaud Perrier issues an open call to developers: “We want you to try it with your media, your application in your environment.” - Code repositories, pipeline definitions, and pre-integrated MCP tools are accessible publicly via Intel's GitHub repositories for community evaluation.

## Notable quotes > [21:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1266s) The bottom line is to start CPU first to simplify your stack on a unified footprint which offers you a you know in excess of six times lower token costs compared with you know typical cloud GPU instances. — Arnaud Perrier

## Mentioned Alliance for Open Media (AOMedia), Arnaud Perrier, Bluedot, Broadpeak, Dell PowerEdge, Docker Compose, FFmpeg, FlashVSR, Intel, Intel Advanced Matrix Extensions (Intel AMX), Intel Agentic Toolkit, Intel Media AI Framework, Intel Quick Sync Video, Intel Video Super Resolution (IVSR), Intel Xeon, Intel Xeon 6, Kubernetes, Media Transport Library (MTL), Model Context Protocol (MCP), NVIDIA RTX 6000, OpenAI Whisper, OpenVINO, Opus-MT, PyTorch, Ray, Richard Chuang, SGLang, Small Pixels, SVT-AV1, Unified Video Quality (UVQ), vLLM, VMAF, Wan 2.1, Wan 2.2

## Video description Deploying AI video at scale often introduces severe infrastructure bottlenecks and cost challenges, driven primarily by heavy reliance on GPUs alone, leaving CPU cycles underutilized, even as AI is rapidly expanding across all tiers of infrastructure.

By leveraging latest-generation Intel® Xeon® processors with Intel® Advanced Matrix Extensions (Intel® AMX), organizations can rebalance their AI workloads by executing compute-intensive workloads such as Video Super-Resolution, inverse tone mapping, and bitrate optimization directly on the CPU alongside encoding pipelines.

Building on this efficient compute foundation, the Intel® Media AI Framework transforms raw media into trusted, searchable AI-ready intelligence, while simplifying deployment of media workloads. Through agentic orchestration, intelligent model routing, reusable MCP tools, and video RAG, the framework powers configurable workflows like semantic search, summarization, metadata enrichment, contextual ad placement, and other MCP-based agent flows.

This session demonstrates the larger value of bringing Xeon-based acceleration through the Intel® Media AI Framework to improve user experience, convert media into actionable intelligence, and empower customers to build and scale cost effective media AI pipelines optimized for Intel® Xeon®.

Presenters: Arnaud Perrier and Richard Chuang from Intel

[Source](https://www.youtube.com/watch?v=SDHzLaOvC0o)

中文简报

英特尔力推至强替代GPU，将视频AI成本削减至六分之一直接在至强处理器上运行智能体媒体管线，通过消除串行工具调用和状态管理中的PCIe总线瓶颈，相比云端GPU可将Token成本降低六倍以上。本次会议探讨了纯GPU视频管线的运行瓶颈，并提出了一种以CPU为先的架构模型，用于扩展媒体画质增强与AI推理。英特尔展示了如何将视频超分辨率（VSR）、SVT-AV1感知码率优化以及Whisper音频转录等任务卸载至英特尔至强CPU，从而显著降低功耗与云端Token成本，并释放GPU算力。为了将原始像素处理与更高层级的语义理解相结合，主讲人详细介绍了英特尔媒体AI框架（Intel Media AI Framework），该框架是一个包含可视化管线构建、动态推理路由及优化MCP工具的开放模块化生态。具体的实际落地案例验证了该方案的可行性，包括动态上下文广告插入，以及在单台至强服务器上以远超实时广播的速度运行的实时多模态篮球分析管线。英特尔展示了运行在搭载Intel AMX的英特尔至强CPU上的高吞吐量视频超分辨率（VSR），1080p转4K超分速度超过900 fps，且功耗相比等效GPU部署降低了80%。 (3:06) 通过在SVT-AV1编码前加入经OpenVINO和AMX优化的AI感知预处理模型，英特尔仅使用额外四个CPU核心便实现了24%至28%的码率降低。 (8:08) 全新的英特尔媒体AI框架整合了模型上下文协议（MCP）工具、智能体工具套件和可视化管线设计器，将原始像素增强与语义理解及动态广告插入统一结合。 (12:31) 为智能体媒体智能采用CPU为先的架构消除了PCIe延迟瓶颈，利用大容量系统内存扩展智能体状态历史记录，与典型的云端GPU实例相比，Token成本降低至六分之一以下。 (21:06)

Reading summary

Intel engineers Arnaud Perrier and Richard Chuang argue that modern media delivery requires moving beyond brute-force encoding toward a unified, CPU-first architecture that couples pixel-level AI enhancement with real-time agentic reasoning. Pixel Enhancement and Generative Offloading (3:06) Traditional transcoding pipelines are transitioning into AI-assisted enhancement workflows that handle Video Super-Resolution (VSR), frame rate conversion, and inverse tone mapping. In benchmarks conducted with Small Pixels, an Intel Xeon CPU utilizing Intel Advanced Matrix Extensions (AMX) delivered over 900 frames per second for 1080p-to-4K upscaling, achieving parity with native 4K output while reducing power consumption by 80% compared to typical GPU configurations. In generative AI video pipelines, offloading super-resolution to the CPU relieves severe GPU bottlenecks. Generating a 5-second 1080p video clip natively using models like Wan 2.1 or 2.2 requires roughly 36 minutes on an NVIDIA RTX 6000 GPU. By generating the initial video at 480p in 5 minutes and handing off FlashVSR upscaling to a Xeon CPU, developers achieve an estimated 7x faster total workflow while freeing GPU memory for parallel clip generation. Perceptual Encoding and High-Throughput Audio Inference (8:08) High-efficiency video compression typically demands non-linear compute trade-offs; switching SVT-AV1 from preset M5 to M1 yields an 18% bitrate reduction but imposes a 10x CPU and latency penalty. To bypass this barrier, Intel introduced an AI-driven perceptual pre-processing model optimized with OpenVINO and AMX. Perrier states that running this pre-processor before encoding cuts bitrates by 24% over M1 at equivalent complexity using only four additional CPU cores, while deep integration reduces bitrates by up to 28%. Partner ISV Bluedot deployed this approach to achieve average 20% bitrate savings on real-time 4K broadcast video without visual degradation. Simultaneously, host CPUs handle high-throughput audio workloads without discrete accelerators. A dual-socket server equipped with Intel Xeon 6 AP processors running OpenAI Whisper Large v3 achieves over 2,000 tokens per second, transcribing more than 600 hours of audio within a single clock hour to support automated subtitling and multi-language translation via Broadpeak. Modular Orchestration with Media MCP Frameworks (12:31) Legacy media workflows frequently rely on hardcoded pipelines that lack the flexibility required for semantic understanding and search. To address this, Intel created the Media AI Framework, a four-tier architecture consisting of a visual pipeline designer, Model Context Protocol (MCP) tool integrations, the Intel Agentic Toolkit, and underlying Xeon silicon. The platform supplies pre-optimized models through a dedicated model zoo to eliminate manual inference tuning. For operational management, the framework incorporates dynamic inference routing to dispatch queries to either compact or large models depending on prompt complexity, balancing latency and token spend. Development can be scaffolded from local Docker Compose environments up to multi-node Ray and Kubernetes clusters, accompanied by guardrails and token-cost tracking. The Architectural Case for CPU-First Agentic Workflows (21:06) Perrier contends that agentic AI fundamentally differs from batch matrix multiplication because it is dominated by continuous loops of serial logic, API gateway calls, state machine updates, and tool coordination. Running both orchestration and inference on Intel Xeon eliminates PCIe data transfer bottlenecks between host processors and accelerators, while standard system memory comfortably holds expanding multi-modal agent histories that would exhaust GPU VRAM. Based on this unified footprint, Perrier claims token costs can be over six times lower than typical cloud GPU setups. Demonstrating this capability, a single Xeon 6 server processed an uncompressed 1080p 30-fps basketball feed, executing parallel computer vision shot detection and live commentary transcription in real time while utilizing only a single-digit percentage of the CPU's total compute capacity. Intel asserts that unified CPU execution provides the operational latency, memory capacity, and cost profile necessary to sustain high-throughput agentic media pipelines. However, realizing these benchmarks requires developers to adopt Intel-specific acceleration libraries, AMX instructions, and modular MCP frameworks across their existing transcoding and inference infrastructure.

精炼阅读摘要

英特尔专家Arnaud Perrier与Richard Chuang指出，应对生成式AI带来的流媒体激增，必须从单纯的像素处理转向结合实时推理与AI的统一架构；利用配备AMX的Intel Xeon处理器作为核心计算底座，能够在降低功耗与延迟的同时兼顾智能编排。像素级处理优化与CPU超分辨率加速 (3:06) 现代视频管线正在向集成了视频超分辨率（VSR）、帧率转换及逆色调映射（ITM）的AI管线演进。针对算力瓶颈，英特尔与Small Pixels合作，在支持Intel AMX的Intel Xeon处理器上运行优化后的VSR模型，实现1080p至4K超分辨率处理速度超900 fps，达到原生4K画质，并比同等GPU部署降低80%的功耗。在生成式AI视频工作流中，CPU与GPU协同展现出显著效率。例如使用Wan 2.1/2.2模型在RTX 6000 GPU上直接生成5秒1080p视频需耗时约36分钟；改由GPU生成480p（耗时5分钟）并交由Xeon CPU运行FlashVSR超分放大后，端到端工作流提速约7倍，同时释放了GPU显存以支持多剪辑并行生成。感知编码压缩与高通量音频推理 (8:08) 在传统编码中，SVT-AV1从平衡预设M5切换至高压缩预设M1虽可节省18%码率，但会导致CPU负载与编码延迟激增10倍。英特尔通过OpenVINO和AMX优化的AI感知预处理模型在编码前剔除不可见细节，仅多消耗4个CPU核心即在相同复杂度下比M1降低24%码率；深入集成至SVT-AV1循环后，码率降幅最高可达28%（对比M1）和25%（对比M5）。合作伙伴Bluedot在其优化器中应用该技术，在4K实时流上实现了20%的码率节约且无肉眼画质损失。音频推理方面，Xeon处理器可在无独立加速卡环境下高效承载语音任务。双路Dell PowerEdge服务器搭载Intel Xeon 6 AP处理器运行Whisper Large v3，推理吞吐量超过2,000 tokens/s，单时钟小时内可转录超600小时音频。Broadpeak已将该方案整合至其软件中，用于实时多语言字幕翻译。模块化媒体AI框架与智能体编排 (12:31) 针对传统媒体流水线硬编码、难以复用的弊端，英特尔推出了开源模块化的Media AI Framework。该架构分为四层：顶层为控制面与可视化管线设计器；第二层是针对英特尔平台优化的MCP（Model Context Protocol）工具集与API网关；第三层包含运行时、模型库及底层执行引擎的Intel Agentic Toolkit；底层则是Xeon处理器硬件基础。该框架的可视化设计器允许开发者通过拖拽验证过的组件快速原型化多模态视频管线，灵活绑定关键帧提取器、语音模型和视觉语言模块。针对生产环境，Agentic Toolkit内置了动态推理路由功能，根据提示词复杂度将请求分发至轻量或大型语言模型以降低成本，并支持每百万Token成本监控及从Docker Compose到Kubernetes与Ray的多节点弹性扩展。情境广告插入与CPU优先架构优势 (18:09) 在商业应用方面，该框架展示了自动化、情境化的视频广告动态插入方案。系统首先抓取饮料品牌的公开网页素材自动生成L型横幅，再利用关键帧视觉分析与音频推理进行全片语义检索，精确定位剧情契合的时间窗口并无缝嵌入广告，兼顾了线性广播与非线性流媒体的分发需求。 Arnaud Perrier强调，智能体AI并非单纯的批量文本生成，而是包含大量串行分支逻辑、API调用、状态维护和工具执行的持续循环。相较于在CPU与GPU间来回传输数据带来的PCIe带宽与延迟开销，统一在Xeon上运行不仅消除了数据搬运瓶颈，还能发挥高单线程性能来解析JSON和执行调度脚本，并借助大容量系统内存承载膨胀的智能体状态数据。Perrier称该架构相比典型云端GPU实例可带来超过6倍的Token成本降幅。在高中篮球比赛直播演示中，单颗Intel Xeon 6处理器同时跑通1080p30实时视频逐帧投篮分析与同步音频转录，计算开销仅占该CPU极小比例。英特尔已将相关代码、管线配置与MCP工具开源至GitHub供行业评估，但该方案在更复杂的超大规模商用场景下的综合表现仍有待进一步落地检验。

Detailed notes — zh

## Background 该视频由 YouTube 频道 @Scale 于 2026 年 11 月 2 日发布，主讲人为来自英特尔（Intel）的 Arnaud Perrier 与 Richard Chuang。视频简介阐述了在部署大规模 AI 视频时过度依赖 GPU 所造成的架构瓶颈与高昂成本，并介绍了如何利用搭载 Intel AMX 的至强（Intel® Xeon®）处理器和 Intel® Media AI Framework，将密集型媒体处理与 Agentic 推理工作流高效卸载至 CPU 执行。 ## TL;DR 本次演讲探讨了媒体处理从传统硬件编解码演进到一体化 AI 与智能体（Agentic）推理的完整路径。英特尔展示了通过 CPU 运行超分辨率（IVSR）、感知预处理优化 SVT-AV1 编码及高吞吐 Whisper 转录，从而显著降低能耗并节约最高 28% 码率的技术方案。随后，演讲重点剖析了四层架构的 Intel Media AI Framework，详细演示了结合 MCP 工具的可视化流水线编排、上下文原生广告植入，以及单路至强服务器上低资源占用、高并发的实时多模态视音频分析与高光切片能力。 ## Atmosphere & style 整场演讲呈现出典型工业级技术发布会的务实与严谨基调，两位主讲人配合默契、论述层层递进。Arnaud 稳健笃定地拆解底层媒体算力与成本效益，Richard 则通过流畅的架构拆解与原型演示将抽象的 Agentic 框架具体化。现场展示的实时多路篮球赛分析与极低的 CPU 占用率构成了极具说服力的技术高潮，让整场分享充满硬核工程感与可落地的架构自信。 ## Bulletin - [12:31](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=751s) Intel 推出 Intel Media AI Framework，通过可复用 MCP 工具、智能路由与流水线设计器，将底层媒体编解码与上层智能体推理统一于 Intel Xeon 处理器。 - [03:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=186s) 通过在 Xeon CPU 上执行 Flash VSR 超分辨率，文生视频模型 Wan 2.1 的渲染流水线加速约 7 倍，同时释放 GPU 显存用于并行生成任务。 - [08:08](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=488s) 基于 AMX 的感知预处理模型仅需 4 个 CPU 核心，即可让 SVT-AV1 在相同复杂度下降低最高 28% 码率，大幅优于单纯调节编码预设的效果。 - [21:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1266s) 针对包含串行逻辑与状态管理的智能体工作负载，采用 CPU 优先架构可消除 PCIe 通信延迟并利用海量系统内存，将推理 Token 成本降低多达 6 倍。 ## Key takeaways - [00:05](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=5s) 媒体处理架构正在经历从传统硬件编码加速（Quick Sync）、高级软件压缩（SVT-AV1）到融入 AI 及 Agentic 实时推理理解的代际演进。 - [03:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=186s) 与 Small Pixels 合作在 Intel Xeon 上运行超分辨率（IVSR），实现 1080p 至 4K 突破 900+ FPS 的吞吐，比典型 GPU 平台节省约 80% 能耗。 - [03:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=186s) 文生视频等 GenAI 任务可通过“GPU 低清渲染 + CPU 超分放大”的异构分流方案，将生成工作流耗时缩短至约七分之一。 - [08:08](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=488s) AI 感知预处理模型结合 OpenVINO 与 AMX 矩阵扩展，能在极低 CPU 开销下显著打破 AV1 编码在压缩率与计算复杂度之间的传统权衡瓶颈。 - [08:08](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=488s) 双路 Intel Xeon 6 服务器运行 Whisper Large V3 音频转录模型，吞吐量突破 2,000 tokens/秒，每小时可处理超 600 小时音频内容。 - [13:29](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=809s) Intel Media AI Framework 采用控制平面、MCP 工具网关、智能体工具包与硬件底座的四层解耦架构，解决媒体开发中的硬编码与孤岛问题。 - [15:35](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=935s) 流水线设计器（Pipeline Designer）支持低代码拖拉拽编排，并提供根据输入复杂度动态分流的智能推理路由机制。 - [17:57](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1077s) 结合多模态视频理解与 MCP 工具，系统可自动解析商品特征并在最契合的时间戳无缝植入上下文原生广告。 - [21:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1266s) 智能体系统重度依赖串行逻辑、API 调度与状态维护，CPU 在单线程响应和海量内存扩展上的优势使其相较 GPU 方案展现出 6 倍的 Token 成本优势。 - [22:23](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1343s) 单路 Xeon 服务器即可在个位数负载开销下，完全实时并发运行 1080p/30 FPS 的比赛高光生成与同步语音转录双流水线。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=5s) 媒体 AI 的规模化挑战与发展历程媒体行业正经历视频流数量的爆发式增长，尤其是生成式 AI（GenAI）模型驱动的实时流媒体，其在各大平台上的可用性正日益提高。规模化交付这些视频流变得极其具有挑战性，需要一个融入 AI 的高效率处理流水线。更重要的是，平台需要具备实时的“深度媒体理解与推理能力”（deep media understanding with reasoning in real time），以此增强用户体验，并进一步实现个性化推荐与商业变现。英特尔媒体团队的 Arnaud Perrier 与 Richard Chuang 共同探讨英特尔及其合作伙伴如何通过基于 CPU 的高能效推理以及专为媒体设计的 Agentic 架构，来加速和简化这一转型过程。回溯英特尔在媒体领域的研发路径，团队长期致力于解决行业遭遇的各种性能瓶颈。最早是通过硬件编码加速，例如在 CPU 中集成 Quick Sync Video 来实现视频流的扩展；随后推出了先进的软件压缩技术——开源编解码器 SVT-AV1（由开放媒体联盟 AOM 打造），该技术甚至荣获了技术艾美奖（Technology & Engineering Emmy Award）。然而，当开发者开始在编解码管线旁边并行引入神经网络与 AI 模型时，整个处理流水线变得支离破碎，开发者陷入了各自为政的孤岛状态。为了打破孤岛，英特尔开源了用于媒体画质增强的 IVSR 库（Intel Video Super Resolution），提供超分辨率等功能，后续又开源了用于专业视频制作环境中未压缩原始视频可靠 IP 传输的 Media Transport Library（MTL）。面对新一轮的技术跃迁，Arnaud 指出当下的核心命题已变成：“开发人员和营销人员现在如何才能进一步提升用户体验并在更大规模上实现商业变现？这确实需要将融入 AI 的媒体流水线与具有可操作性的媒体理解能力结合起来。” ```mermaid flowchart LR accTitle: 媒体处理演进路径 accDescr: 展示从硬件编码加速到融入 AI 与 Agentic 架构的四个演进阶段 subgraph S1["阶段 1：硬件加速"] A["Quick Sync Video 编码"] end subgraph S2["阶段 2：高级软件压缩"] B["SVT-AV1 编解码器 (艾美奖获奖项目)"] end subgraph S3["阶段 3：独立 AI 与传输库"] C["IVSR 库 (超分辨率等增强)"] D["Media Transport Library (IP 原始视频传输)"] end subgraph S4["阶段 4：一体化媒体 AI"] E["融入 AI 的管线 + Agentic 推理理解"] end A --> B --> C & D --> E ``` --- ### [03:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=186s) AI 融入的媒体流水线与三大核心增强模型标准视频编解码处理工作流正被 AI 深度重构，跨越传统转码阶段，迈向智能画质增强与用户体验升级。AI 在工作流的各个环节均带来变革： 1. **视频层面**：通过视频超分辨率（VSR）增强画质，利用帧率转换（FRC）实现更流畅的动态效果，借助反向色调映射（ITM）提升画质并完整保留细节。 2. **音频层面**：自动语音识别（ASR）与文本转语音（TTS）为内容赋予元数据价值和语义丰富度。 3. **质检评估**：画质评估指标也在随着 AI 演进，如 VMAF 以及最新的 UVQ（Unified Video Quality）指标，使画面增强效果能够被全自动、可量化地衡量。这些 AI 插件依托原生生态与 PyTorch、SGLang、vLLM 以及英特尔 OpenVINO 等主流框架兼容，并通过英特尔开源框架缝合进 FFmpeg AI 插件中，以确保极高的灵活性与开箱即用的易用性。 #### 1. 视频超分辨率（VSR / VSSR）与能效优化超分辨率技术被广泛应用于广播、流媒体、UGC 内容、广告以及 FAST 频道（免费流式广告电视），尤其用于提升付费订阅会员的高端画质体验。由于没有任何单一的 VSR 模型可以满足所有业务场景对“质量 vs 性能”的权衡要求，英特尔在 IVSR 库中提供了丰富的合作伙伴算法及经过优化的开源预训练模型。 - **与 Small Pixels 合作案例**：基于英特尔至强（Intel Xeon）CPU 运行超分模型，在将视频从 1080p 超分放大至 4K 分辨率时，帧率突破 **900+ FPS**。 - **画质与能耗**：放大效果逼近原生 4K 画质，同时相较于典型的 GPU 方案，“节省了约 80% 的能耗”（saving about 80% of the energy footprint you would consume on a typical GPU platform），显著压缩了运营成本与总耗电量。 #### 2. VSR 辅助生成式 AI（GenAI）内容生成随着生成式视频模型的爆火，端到端高分辨率渲染面临严重的算力瓶颈。英特尔通过开源文生视频模型 Wan 2.1（原识别为 1 2.2）测试，从提示词生成一段极其细节丰富的 5 秒 1080p 视频。 - **GPU 渲染代价**：在高端专业 GPU（RTX 6000）上，生成这段 5 秒 1080p 片段耗时长达 **36 分钟**。这在面向数百万并发用户的规模化生产中极为昂贵且不切实际。 - **低清生成 + CPU 超分方案**：若将 GPU 渲染分辨率降至 480p，生成时间可压缩至 **5 分钟**，但画面模糊会导致用户流失。工程团队引入高质量的 Flash VSR 模型，在 CPU 上把 480p 画质拉伸回 1080p，画面细节非常接近原始高画质渲染。 - **架构解耦价值**：在至强 CPU 上运行 Flash VSR 释放了宝贵的 GPU 显存，使 GPU 能并行渲染多个片段或更长的片段，将 GPU 彻底变为工作流加速器，整体工作流速度提升了约 **7 倍**，兼顾了吞吐效率与画质。 ```mermaid flowchart TD accTitle: VSR 辅助 GenAI 视频生成管线 accDescr: 展示降低 GPU 渲染分辨率并在 CPU 上执行 Flash VSR 带来的并行效率与加速 subgraph GPU["GPU 任务池 (显存释放)"] P["Prompt 提示词"] --> G1["Wan 2.1 生成 (480p 低清)"] P -.->|"并行释放显存"| G2["并行渲染更多/更长切片"] end subgraph CPU["Intel Xeon CPU 卸载"] G1 -->|"480p 视频流"| C1["Flash VSR 超分辨率增强"] C1 --> OUT["1080p 高画质交付 (耗时缩短至 ~1/7)"] end ``` --- ### [08:08](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=488s) 感知预处理优化 AV1 编码与高吞吐音频转录在高质量软件视频编码中，开发者长期面对“压缩率 vs 计算复杂度”的权衡博弈。为了在超热门视频的分发中降低码率、节约 CDN 带宽，视频工程师常常需要付出指数级的计算资源与 CPU 周期。以社交媒体和流媒体中被广泛采用的 SVT-AV1 为例： - 从标准操作预设 M5 调整到极端高质量预设 M1，虽然能减少约 18% 的码率，但编码速度慢了将近 10 倍，需要消耗 **10 倍以上的 CPU 资源**。为了突破这一瓶颈，英特尔设计了一种基于 OpenVINO 训练并借助至强 AMX（Advanced Matrix Extensions，高级矩阵扩展）指令集加速的 AI 感知预处理模型（Perceptual Pre-processing Model），在视频进入编码器之前先进行画质预处理。 - **编码突破**：仅在 CPU 上额外使用 **4 个核心**，该模型就能在相同复杂度下使 SVT-AV1 实现比纯 M1 预设还要低 **24% 的码率**，Arnaud 评价道：“在相同的复杂度下，比 M1 的码率降低了 24%，无论采用何种预设，单纯依靠 AV1 编码技术或编解码器本身根本不可能做到这一点。” - **深度整合与量化收益**：在多模型架构下根据量化参数（QP）联合调优，该方案相比单独运行 M1 预设实现了最高 **28% 的码率降低**；相比基准 M5 预设实现了 **25% 的码率节省**。 - **BlueDot ISV 合作优化**：合作伙伴 BlueDot 进一步将此类算法优化为其专有的“Perceptual Quality Optimizer”，同样作为预处理模块实时运行在至强处理器上。结合 OpenVINO 与 AMX 加速，在 4K 超高清视频编码中实现了稳定的平均 **20% 码率节省**，且肉眼察觉不到任何画质损失。在音频流处理方面，为了帮助媒体平台实现全球内容分发和自动化字幕变现，流水线集成了语音识别与机器翻译： - 集成开源 Whisper 语音识别模型（Speech-to-Text）与 Opus-MT 翻译模型（Language Translation）。 - **单路服务器性能**：单颗 Xeon 6 SoC 即可实时运行 2 路完整的实时转录翻译流。 - **高密集群吞吐**：在双路 Dell PowerEdge 等级的高性能服务器上搭载 Intel Xeon 6 处理平台时，运行 Whisper Large V3 模型的吞吐量**超过 2,000 tokens/秒**。 - **转录能力换算**：这相当于系统**每小时可以处理超过 600 小时的音频内容转录**，保持基准参考级别的准确率，且全过程完全脱离高功耗 GPU。 - 合作伙伴 Broadpeak 已将该方案整合至其商业软件产品中，实现了动态多语言字幕实时切换与生成。 --- ### [12:31](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=751s) 从硬编码流水线跃升至 Agentic 媒体框架随着音频与视频端到端 AI 插件的成熟，Arnaud 将演讲移交给同事 Richard Chuang，共同展开英特尔媒体 AI 布局的下一个战略飞跃：将“底层媒体处理”与“高层 AI 推理”（Media Processing and AI Reasoning）有机统一。以往传统的媒体 AI 业务大多针对特定单一需求编写“硬编码”（hardcoded）逻辑，导致系统僵化且难以升级复用。针对这一痛点，英特尔媒体 AI 框架（Intel Media AI Framework）旨在提供一系列经过验证的、可复用的组件，以极大缩减开发者的开发周期，解决当前行业面临的三大鸿沟： 1. **工具与模型可复用性（Reusability）**：打破一次性代码，使所有基础工具、推理插件和模型接口实现标准化、模块化封装。 2. **预先极致优化的模型库（Optimized Model Zoo）**：模型部署优化通常需要占用开发者海量的时间与精力，英特尔在自建的 Model Zoo 中直接交付适配好 Xeon CPU 与 AMX 指令集的高性能优化模型，开发者可以直接开箱即用，免去了繁琐的底层硬件调优。 3. **连接复杂媒体与智能推理**：为引入 Agentic 框架打下基础设施基石，让媒体流水线从被动的画质与编码调优升级为具备自主感知、实时分析与推理的智能体系统。 ### [13:29](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=809s) Intel 媒体 AI 框架的四层解构与 MCP 工具集成 Richard Chuang 阐述了 Intel 媒体 AI 框架（Intel Media AI Framework）在模型上下文协议（MCP, Model Context Protocol）方面的集成能力。针对开发者面临的工具繁杂、自研适配成本高的问题，Intel 直接提供了一系列开箱即用的 MCP 工具，开发者无需从零构建各类工具接口，更重要的是“这些 MCP 工具均已针对 Intel 平台进行了深度优化”。随后，Richard 将该框架自顶向下拆解为清晰的四层架构体系： ```mermaid flowchart TD accTitle: Intel 媒体 AI 框架四层架构 accDescr: 展现从上层控制平面到硬件基础设施的四层解耦架构与开发者接入方式 subgraph Layer1["第 1 层：控制平面与流水线设计器"] CP["Control Plane & Pipeline Designer (智能体工作流编辑 / 注册表 / 访问控制)"] end subgraph Layer2["第 2 层：MCP 工具与 API 网关"] MCP["Media AI MCP Tools & API Gateway (Intel 平台优化工具集)"] end subgraph Layer3["第 3 层：Intel 智能体工具包"] ATK["Intel Agentic Toolkit (优化库 / 模型库 / 运行环境 / 仓库)"] end subgraph Layer4["第 4 层：硬件底座"] HW["Intel Xeon 处理器平台 (高效算力执行底座)"] end DevApp["开发者应用 / 终端交互"] -.->|"访问控制与 UI"| Layer1 MCPApp["MCP 应用 / Agent"] -.->|"直接调用工具集"| Layer2 DeepDev["深度定制开发者"] -.->|"完全控制底层组件"| Layer3 Layer1 --> Layer2 Layer2 --> Layer3 Layer3 --> Layer4 ``` - **第一层（控制平面与流水线设计器）**：提供可视化界面用于编辑智能体工作流（Agentic Workflow），内置模型注册表（Registry）与访问控制（Access Control）。 - **第二层（媒体 AI MCP 工具与 API 网关）**：封装了经过底层性能优化的 MCP 工具集与服务网关，支持外部应用直接调用。 - **第三层（Intel 智能体工具包，Intel Agentic Toolkit）**：提供高度优化的底层加速库、主流模型资产库（Model Zoo）、运行时环境及组件仓库，供需要全面控制权的开发者使用。 - **第四层（硬件底座）**：基于 Intel Xeon 处理器构建，为全套软件堆栈提供高能效的算力支撑。面向开发者生态，框架在右侧开放了多元化的接入方式：终端应用可以通过预置的图形界面或 API 调用接入整个框架；具备 MCP 交互能力的应用可直接与第二层的 MCP 工具集握手；而希望获得完全掌控权的开发者，则能够直接穿透到底层的 Intel 智能体工具包，实现从开发、调试到部署的全流程自主定制。 ### [15:35](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=935s) 可视化流水线设计器与全功能智能体工具包为了消除智能体开发的复杂门槛，框架推出了流水线设计器（Pipeline Designer）。开发者无需编写繁琐的定制胶水代码，即可通过图形化托拉拽方式直观编排工作流，利用经官方验证的组件快速组装原型。应用若需联动特定的 MCP 工具并调用预置模型，开发者只需在该界面中将对应的 MCP 工具与已注册的模型节点进行绑定，大幅简化了端到端的技术落地路径。针对支撑框架运转的底层工具链，Richard 特别指出：“智能体所需要的远不止推理模型、MCP 服务器和数据库。”完整的架构必须具备强大的外围支撑能力，以便开发者免于陷入漫长琐碎的适配与调优工作。该工具包的核心支撑功能涵盖： - **智能推理路由（Inference Routing）**：能够根据输入 Prompt 的复杂度特征，动态将请求分流至轻量级模型或大规模语言模型，平衡响应延迟与生成质量。 - **高性能推理与服务运行时**：支撑各功能模块在统一底座上平稳承载并发推理。 - **生产级运维管控机制**：内置访问控制、安全护栏（Guardrails）与全链路可观测性，帮助运维团队实时监控每百万 Token 支出成本（Cost per Million Tokens）。 - **弹性部署策略**：支持极简的单节点 Docker Compose 快速验证，并可平滑横向扩展至基于 Kubernetes 与 Ray 的多节点分布式计算集群。 ### [17:57](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1077s) 实战落地：基于视频理解的上下文原生广告植入 Richard 通过“电影内上下文果汁广告植入”的实际商业场景，展示了该媒体 AI 框架的落地能力。传统广告插播往往生硬割裂，而该方案首先使用框架自带的关键帧视觉模型与音频推理工具对影片内容进行深度多模态分析。 ```mermaid flowchart LR accTitle: 智能上下文广告插入端到端流程 accDescr: 描述从果汁商品理解到视频语义匹配与最终 L 型横幅展示的完整流程 A["分析公共网页 理解果汁产品特征"] --> B["自动生成品牌 L 型广告横幅"] C["影片多模态深度理解 (视觉关键帧 + 音频推理)"] --> D["视频对视频语义检索 锁定自然契合时间点"] B --> E["结合时间戳与物料"] D --> E E --> F["移交 Prophesee 解决方案 执行动态广告合成"] F --> G["终端观众视效 (自然融入，无打扰)"] ``` 框架通过公开网页抓取解析并理解果汁商品卖点，自动生成对应的 L 型广告横幅（L-shaped banner）；同时，利用视频对视频语义检索，精准定位影片中与饮品消费场景最契合的情节瞬间并提取时间戳。随后，这一时间戳被移交至下游的 Prophesee 广告植入方案中完成动态渲染，既支持线性流媒体也兼容非线性视频场景。这种深度契合剧情的无缝呈现方式，实现了广告主营销转化、内容版权方商业变现与订阅用户观影体验三方共赢的生态循环。 ### [19:30](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1170s) 闭环开发旅程：从零代码探索到分布式规模化部署该框架为开发者规划了一条“探索—构建—规模化”的渐进式路径。开发者最初进入控制界面，在不需要编写任何代码的前提下，即可直接试用各类多模态 AI 与智能体工具包；确认可行性后，再借助流水线设计器拖入并绑定视觉关键帧、语音识别及大语言模型，启动本地端到端测试。在调试过程中，若输出质量不及预期，开发者可随时暂停运行并返回流水线设计器，按需替换绑定的 MCP 工具或切换模型版本，重新运行即可快速验证。“因此，迭代循环可以非常高效且迅速地运转，从而达成预期的目标产出。”当原型逻辑在单节点验证成熟后，开发者即可一键将该流水线推送至 Kubernetes 或 Ray 分布式集群进行大规模生产级扩展，平滑覆盖从初次技术验证到大规模商用部署的完整生命周期。 ### [21:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1266s) CPU 算力架构优势与六倍 Token 成本优化 Intel 的 Arnaud Perrier 接过讲解，总结了智能体计算的系统级本质。他强调指出： > “智能体 AI 是一个由串行逻辑、API 调用和状态管理构成的连续循环，而绝不仅仅是简单的文本生成。” > —— Arnaud Perrier 在这种工作负载特征下，传统单纯依赖外挂加速卡的方案面临严重瓶颈，而 CPU 展现出不可替代的体系结构优势： 1. **统一计算底座，消除数据搬移延迟**：在 CPU 上同时承载控制平面与模型推理，免去了主机内存与外置 GPU 显存之间频繁通过 PCIe 总线拷入、拷出数据的严重通信延迟。 2. **高频单线程推理与编排调度性能**：智能体在处理复杂的 JSON 反序列化、执行 Python 逻辑脚本或调度工具 API 时，严重依赖 CPU 的单线程响应速度。 3. **海量系统内存扩展空间**：快速膨胀的多轮交互智能体上下文状态，能够直接留存在容量庞大且成本可控的系统 RAM 中，彻底摆脱了加速卡显存不足的硬件墙。基于此，Arnaud 提出了“CPU 优先（Start CPU first）”的基础架构选型原则。将技术栈精简收敛于统一部署架构上，相较于典型的云端 GPU 实例，能够将推理成本降低 6 倍以上（6x lower token costs）；当业务场景确实需要运行超大规模模型时，该架构依然保留了平滑向上扩展至独立 GPU 的弹性空间。 ### [22:23](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1343s) 现场实测演示：单路 Xeon 上的多路并行实时视音频分析 Arnaud 在演讲尾声展示了一段高拟真度的实时视觉多模态理解演示。现场采集了一场高中篮球赛的实时摄像机信号，流媒体直接输入至一台单路 Intel Xeon 处理器服务器中进行全实时分析。同一颗 CPU 上并行并发运行着两个独立的 AI 处理流水线： - **视频处理流水线**：实时对 1080p、30 FPS 的全画幅视频进行逐帧特征计算，精准捕获每一次投篮出手的精彩瞬间，并实时剪辑生成高光片段卡片（Highlight Clip）。 - **音频处理流水线**：实时捕捉解说音轨，同步转化为流式文字对白显示在视频侧边栏。监控面板显示，两套复杂的视音频模型并发执行时，整体分析速率比真实播放速率快数倍，在全实时吞吐场景下仅占用了该单台服务器个位数的资源百分比（a few percent）。Arnaud 总结表示，这些功能代码均已开源在 GitHub 仓库中，呼吁开发者直接拉取代码仓库，在自己的业务场景与媒体数据上体验 Intel 媒体 AI 框架的实测效能。 ## Notable quotes > [21:06](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=1266s) 智能体 AI 是一个由串行逻辑、API 调用和状态管理构成的连续循环，而绝不仅仅是简单的文本生成。 — Arnaud Perrier > [15:35](https://www.youtube.com/watch?v=SDHzLaOvC0o&t=935s) 智能体所需要的远不止推理模型、MCP 服务器和数据库。 — Richard Chuang ## Mentioned Arnaud Perrier, Richard Chuang, Intel, Intel Xeon, Intel AMX, Intel Media AI Framework, Quick Sync Video, SVT-AV1, Alliance for Open Media, IVSR, Media Transport Library, PyTorch, SGLang, vLLM, OpenVINO, FFmpeg, VMAF, UVQ, Small Pixels, Wan 2.1, NVIDIA RTX 6000, Flash VSR, BlueDot, Whisper, Opus-MT, Broadpeak, Dell PowerEdge, Model Context Protocol, Prophesee, Docker Compose, Kubernetes, Ray ## Video description Deploying AI video at scale often introduces severe infrastructure bottlenecks and cost challenges, driven primarily by heavy reliance on GPUs alone, leaving CPU cycles underutilized, even as AI is rapidly expanding across all tiers of infrastructure. By leveraging latest-generation Intel® Xeon® processors with Intel® Advanced Matrix Extensions (Intel® AMX), organizations can rebalance their AI workloads by executing compute-intensive workloads such as Video Super-Resolution, inverse tone mapping, and bitrate optimization directly on the CPU alongside encoding pipelines. Building on this efficient compute foundation, the Intel® Media AI Framework transforms raw media into trusted, searchable AI-ready intelligence, while simplifying deployment of media workloads. Through agentic orchestration, intelligent model routing, reusable MCP tools, and video RAG, the framework powers configurable workflows like semantic search, summarization, metadata enrichment, contextual ad placement, and other MCP-based agent flows. This session demonstrates the larger value of bringing Xeon-based acceleration through the Intel® Media AI Framework to improve user experience, convert media into actionable intelligence, and empower customers to build and scale cost effective media AI pipelines optimized for Intel® Xeon®. Presenters: Arnaud Perrier and Richard Chuang from Intel [Source](https://www.youtube.com/watch?v=SDHzLaOvC0o)

Source quotes

[21:06] Arnaud Perrier: 智能体 AI 是一个由串行逻辑、API 调用和状态管理构成的连续循环，而绝不仅仅是简单的文本生成。 [15:35] Richard Chuang: 智能体所需要的远不止推理模型、MCP 服务器和数据库。

Source quotes

[21:06] Arnaud Perrier: The bottom line is to start CPU first to simplify your stack on a unified footprint which offers you a you know in excess of six times lower token costs compared with you know typical cloud GPU instances.

Original transcript

[0:05] Hi everyone, welcome to this session [0:07] about scaling media AI. As you know, the [0:11] media industry as a whole is going [0:12] through an explosion of video streams in [0:16] particular live media streams driven in [0:20] large part by geni models and their ava [0:24] increasing availability through the [0:26] platforms. So successfully delivering [0:29] these streams at scale [0:32] is increasingly challenging and requires [0:35] a much more efficient AI infused [0:38] processing pipeline going forward. But [0:41] even more importantly, a deep media [0:44] understanding with reasoning in real [0:46] time so that platforms can further [0:50] enhance the user experience but also [0:53] further personalize and monetize that [0:56] experience. [0:58] I'm Arno, by the way, and joining me [1:01] today is Richard Chwang. And we'd like [1:04] to highlight some of the solutions the [1:06] media team at Intel and and our partners [1:08] are developing to accelerate and [1:11] facilitate that transition through [1:14] efficient inference on CPU and Agentic [1:18] frameworks designed specifically for [1:20] media. [1:22] But first, let me take you back a step [1:26] in time. [1:27] And looking at the journey that Intel [1:30] and our customers and partners have been [1:32] on for a while now, we've hit certain [1:35] challenges along the way in the media [1:37] space. Starting with performance [1:40] bottlenecks like encode acceleration in [1:43] hardware to be able to scale that with [1:46] quicks sync video in the CPU or later [1:50] advanced softwarebased compression with [1:53] SVTV1 which was you know open sourced an [1:56] open source codec from AOM and received [1:59] an technical Emmy [2:02] and as developers started using neural [2:05] networks and AI models on the side. What [2:08] happened is that the processing pipeline [2:10] itself became fragmented. People work in [2:13] silos. [2:15] So to address that we open sourced what [2:17] called the IVSR library for media [2:20] enhancement including functions like [2:22] super resolution [2:25] and later the media transport library [2:27] for reliable IP transport over I you [2:32] know for for video production [2:35] environments and transport of raw video. [2:39] But the next question really is how can [2:42] developers and marketers further enhance [2:45] their user experience and monetize at a [2:48] greater scale at an even greater scale [2:50] now and that really requires a blend of [2:54] AI infused media pipelines as well as [2:57] actionable understanding. [3:00] So let's take that leap together if you [3:03] will. [3:06] Let's start with the media pipeline [3:08] itself. [3:10] So a standard video and encoding [3:13] processing workflow is shown here. Many [3:16] of you will will know it very well. And [3:19] that workflow is transforming with AI [3:21] and moving beyond the basic transcoding [3:24] to intelligently enhance content quality [3:27] and user experience. [3:29] AI adds video enhancement through VSSR, [3:33] creates smoother motion with things like [3:36] frame rate conversion, [3:39] improves visual quality with ITM while [3:42] preserving details, whereas the audio [3:45] enhancements like automatic speech [3:47] recognition [3:48] and text to speech add value and [3:51] enrichment to the content. And the [3:54] transformation is not only happening in [3:56] the the pre-processing stage, but visual [4:00] quality metrics are evolving with AI as [4:03] well to ensure these new enhancements [4:05] are measurable and automated like VMAF [4:09] or or the most recent UVQ metric. [4:13] So seamless integration of those AI plug [4:16] plugins is enabled with out of the box [4:19] support for the major AI frameworks like [4:23] PyTorch, SG Lang, VLM and the Intel [4:27] input OpenVO of course and Intel [4:30] publishes these models into open-source [4:32] frameworks [4:34] and stitch stitch them into FFMAG AI [4:37] plugins [4:39] so as to ensure flexibility and ease of [4:43] use. [4:45] So we're going to highlight three of [4:47] those enhancement models as use cases [4:50] here. The first is super resolution, [4:53] then bit rate reduction and the [4:55] automatic speech recognition use case [4:57] for audio. So let's dive into those. The [5:01] first one is VSSR. It's used, you know, [5:04] commonly and there's been nonAI models [5:06] of it across broadcasting, streaming, [5:08] UGC and advertisement things like fast [5:11] channels to, you know, try to provide a [5:13] better user experience in premium [5:15] streaming subscriptions in particular. [5:17] But no single VSSR model can serve all [5:20] media pipelines requiring flexibility [5:23] between the quality and the performance [5:25] of the model. Hence there are a range of [5:28] them available through partner as well [5:32] as open- source pre-trained optimized [5:35] models. [5:37] We've worked with you know our partners [5:39] to publish a number of those models in [5:41] the IDSR library [5:44] as shown here. But we wanted to [5:46] highlight one recent collaboration with [5:49] small pixels to show you the power of [5:53] you know IDSR Han Xeon CPUs [5:57] at you know 900 plus frames per second [6:00] for 1080p to 4K upscaling. So as you can [6:04] see we can reach you know very very high [6:07] quality close to original native 4K [6:11] while saving about 80% of the energy [6:15] footprint [6:17] you would consume on a on a typical GPU [6:20] platform to do that. So thereby also [6:23] reducing the cost and total power. [6:28] The second use case we wanted to [6:30] highlight is VSSR assisted content [6:32] generation. [6:34] With the popularity of ever growing [6:37] popularity of Genai models, we use open [6:40] source genai model 1 2.2 to generate [6:43] this 5-second clip from a prompt with as [6:46] much detail as possible. And believe it [6:49] or not, but it takes 36 minutes to [6:51] generate that 5-second clip in 1080p [6:55] resolution on a high powered GPU. [6:58] Typically RTX 6,000 is what we used [7:01] here, which gets very expensive and even [7:04] unpractical when you want, you know, [7:06] deploy these at scale for millions of [7:08] users at the same time. Now one can get [7:12] it down the gen time to about 5 minutes [7:16] by lowering the resolution to 480p for [7:19] example but then the user experience is [7:22] poor and you lose engagement. So [7:24] engineers have started using highquality [7:27] flash VSSR model which is a little bit [7:30] more expensive a VSSR model to upres [7:33] that 480p back to 1080p in a quality [7:37] that's very close to the intended [7:40] original. And running flash VSSR on the [7:44] CPU frees up the GPU memory so that you [7:47] can generate either longer clips or [7:49] multiple clips at the same time in [7:51] parallel on the GPU. So the GPU becomes [7:55] an accelerator for the entire workflow [7:57] achieving you know you know in the order [8:00] of seven times faster generation with [8:03] the quality that you want for your [8:05] users. [8:08] Another common challenge we encounter in [8:11] highquality software video encoding is [8:13] the trade-off between compression and [8:16] complexity. So when you want to reduce [8:20] the bit rate, achieve higher compression [8:23] to save on delivery costs for very high [8:27] popular videos or clips. [8:31] Achieving those lower bit rates as you [8:33] know video developers know very well [8:35] require exponentially more CPU cycles [8:39] or time. So in this example, we we're [8:43] using SVT81, which is widely deployed [8:46] across, you know, social media and [8:48] streaming platforms. [8:50] And so going from a standard operating [8:53] preset of, let's say, M5 in the middle [8:55] of the curve to M1 at the far right to [8:58] achieve about an 18% bit rate reduction [9:02] is close to 10 times slower or requires [9:06] 10 times more CPU resources. [9:10] So what we looked into is if you design [9:14] in a AI based pre-processing perceptual [9:18] pre-processing model trained with open [9:21] veno and optimized on Zeon with [9:24] inference acceleration AMX [9:27] when applied before the encoding stage [9:30] you can achieve a 24% iterate reduction [9:34] over M1 at the same complexity which [9:38] would be impossible [9:39] using AV1 encoding technology or just a [9:43] codec at any preset for that matter and [9:46] that's only using an additional four [9:48] cores on the CPU. So that's super [9:51] efficient. So when integrated in the AV1 [9:55] pipeline the the A1 encodering loop [9:58] itself that model we call SVP m you know [10:02] used in a multimodel fashion based on QP [10:06] achieves up to 28% lower ADR than the [10:11] highest quality preset on A1 the M1 [10:13] encode alone and a 25% reduction over [10:18] you know the M5 operating point. [10:22] So that's that's an impressive result. [10:24] And we've worked with partner ISV called [10:28] Blue Dot who's further optimized that [10:32] kind of model [10:34] through their perceptual quality [10:36] optimizer [10:37] again as a pre-processing block before [10:41] the final encode on Xeon processors in [10:44] real time. [10:46] They've optimized it using Okon Vino, [10:49] but also leveraging AMX acceleration and [10:51] reached, you know, a consistent 20% [10:54] average bit rate savings even for [10:57] highquality 4K videos like this one with [11:00] no perceived, you know, quality [11:03] degradation. [11:05] And the final model I wanted to talk [11:07] about is audio part up to pipeline. So [11:10] to improve global reach and automate [11:14] that portion you know revenue generation [11:17] we can tackle audio as well with models [11:19] open source models like whisper for [11:21] transcription speech to text and opusmt [11:24] for translation for language translation [11:27] that's so and with those models running [11:30] again on CPU we can achieve two [11:33] real-time streams on a single von6s [11:36] or on a high power Dell power edge type [11:39] of server [11:41] with Xeon 6 AP processor dual socket you [11:44] can run whisper large V3 at in excess of [11:50] 2,000 tokens per second. So that [11:52] translates to 600 plus hours of audio [11:55] transcribed per hour at reference [11:58] accuracy and all without the use of [12:01] again of a high powered GPU. [12:05] So we've worked with our partner Brock [12:08] Peak and they've integrated that into [12:11] their software application. And this is, [12:14] you know, some examples of dynamically [12:17] changing the language and the [12:19] transcription for subtitles. [12:22] And with that, let me hand it off to my [12:24] colleague Richard to take us through how [12:27] this all comes together in an agentic [12:30] world. [12:31] >> Thanks, Arnold. Now let's dive into the [12:34] next leap we will make to connect media [12:37] processing and AI reasoning together. [12:40] Traditionally media AI workload they are [12:43] hardcoded for specific purpose. With [12:46] this framework we will provide a number [12:48] of components that are reusable verified [12:51] so that we will accelerate the [12:53] development process of your application. [12:56] This slides highlights three gaps that [12:59] we trying to resolve using media AI [13:02] framework. The first one is to make all [13:05] the tool and models reusable that I just [13:08] mentioned. And the second is the model [13:12] has to be optimized and that takes a lot [13:15] of work you know for all of the [13:16] developers and Intel we will provide [13:19] those optimized model in our model zoo [13:21] so that our developer can just use it [13:24] immediately without having to do their [13:27] own optimization. [13:29] The third one is the MCP integration. We [13:32] are going to provide a number of MCP [13:34] tools that is available for you to [13:37] integrate without having to create you [13:40] know all different type of MCP tool by [13:42] yourself and the best part is this MCP [13:44] tool is also optimized for Intel [13:47] platform. So right now let's look at you [13:50] know how we decompose this framework [13:52] into different layers. [13:57] The Intel Media AI framework organizes [14:00] this capability into four layers. At the [14:04] top, the control plane and the pipeline [14:06] designer provide interfaces for editing [14:10] agentic workflow with registry and [14:13] assess control. The second layer is both [14:16] media AI and CP tools and API gateways. [14:19] Below that that is Intel agentic toolkit [14:22] supplied optimized library models [14:26] runtimes and also repositories that is [14:29] available for our developer to use. The [14:32] hardware foundation offer efficient [14:34] execution on Intel Xeon preform. On the [14:38] right, we offer various interface where [14:41] developer can have their application to [14:44] interface with the end user and also [14:47] their application can use a default [14:50] graphical interface and also API calls [14:52] to access the framework for all the [14:55] function that we just mentioned. [14:57] The MCP capable applications that [15:00] developer are working on they can also [15:03] interface directly to layer 2. That is [15:06] where we provide MCP tool sets. Those [15:10] are optimized for Intel platform and we [15:13] make them available directly to the [15:15] application to use or agent to use. And [15:18] third one is a developer who want the [15:21] full control of their agent development. [15:24] They can connect directly to Intel [15:26] agentic toolkit. So you have full access [15:30] of all of the needed component from [15:32] development to deployment. [15:35] Next we will quickly review the ease of [15:38] use feature that we develop in this [15:40] framework. [15:43] The Intel media AI framework over two [15:47] convenient development features. [15:50] Pipeline designer let you compose your [15:53] workflow visually and the prototype [15:57] using verified components with less [16:00] custom code. for customer or for [16:03] developer whose application will want to [16:07] link directly to MCP tool and also [16:10] leverage the the models that we already [16:12] have in the model zoo. They can also use [16:16] this interface to connect the right MCP [16:19] tool together and also bind the right [16:23] model that they want to use that is [16:25] registered in this framework. So this [16:28] path complement each other to make the [16:31] development a developers journey easier. [16:35] Next we will look at the toolkit that [16:38] support the entire entire framework that [16:42] we just talked about. [16:45] Agents need more than reasoning models, [16:48] NCP servers and databases. This [16:51] architecture brings together the [16:53] supporting capability around Intel Xeon [16:56] while allowing developer to work without [16:59] spending too much time to do the [17:00] integration and optimization by themsel. [17:04] Let's focus on some key features that [17:06] we're providing through this toolkit. [17:08] Inference routing connects request with [17:12] smaller or larger language model based [17:15] on the complexity of the prompt. The [17:17] inference toolkit and also serving [17:20] runtime support model execution across [17:23] these functions. SS control guard rail [17:26] and also observability address the [17:28] operational needs such as you need to [17:31] keep your eyes on the cost per million [17:34] tokens that you're spending. At the [17:36] bottom deployment option range from [17:39] docker composer on one node to [17:42] kubernetics and ray across multiple [17:45] nodes. The key is the comprehensive [17:49] infrastructure support around the agent [17:52] including deployment and also ongoing [17:55] operations. [17:57] Next we will show how Intel media AI [18:00] framework contribute to a real world use [18:03] case that is contextual advertisement [18:06] insertion. [18:09] Here we want to place a juice [18:12] advertisement in the movie. The movie [18:15] already been analyzed by Intel media AI [18:19] framework through the video key frame [18:21] and also audio reasoning tool. First we [18:24] have the framework to understand the [18:27] juice product from the public website [18:30] and then we generate the L-shaped banner [18:32] for insertion and then we use the media [18:36] AI framework to run the video to video [18:38] searching and also identify the best [18:41] timing for us to insert a and once we [18:45] have that this time stamp will be handed [18:48] to props banner in two solution to [18:53] display the at banner at the right [18:55] timing. This is what the subscribers [18:59] will see. The Intel media AI framework [19:03] provides the media understanding while [19:06] propic handle the ad insertion [19:09] supporting both linear and nonlinear [19:11] placement. Together they help the ad fit [19:15] naturally with the content and also [19:17] improve the viewers experience which is [19:20] a win-winwin scenario for advertisers, [19:24] content owner and also the subscribers. [19:30] We envision the media AI framework to [19:33] provide a full development journey [19:35] starting in the control interface. [19:38] Try media AI and also agentic toolkits [19:43] without having to write any code. Then [19:45] use the pipeline designer to compose a [19:49] working flow. And once we start this [19:51] journey, we can have the developer to [19:55] bind the vision, key frame, speech and [19:58] also language model that will be part of [20:01] this agent workflow and have them to [20:04] start running in the environment and [20:06] they will be able to see the results [20:08] quickly. And if they see something that [20:11] need to be changed in order to get to [20:12] better quality output, they can stop it [20:15] and then go back to the pipeline [20:17] designer and change the component like [20:20] MCP tool and the models and then run it [20:22] again. So the iteration loop can run [20:25] very efficiently and also very quickly [20:28] to get to the desired outcome. [20:32] So when ready when the developer ready [20:34] to scale the pipeline they deploy it [20:37] with kubernetics and ray distributed [20:40] computing are always an option. So [20:42] across explore develop and also deploy [20:47] we bring this journey together using the [20:51] media AI framework to support the first [20:54] trial to the production at scale. With [20:58] that, I will hand it over to Arnold to [21:02] recap the key takeaway of today's [21:06] section. Thank you, Richard. So, as a [21:09] summary, Aentic AI is a continuous loop [21:12] of serial logic, API calls, state [21:15] management, and not just simple text [21:18] generation. [21:20] This is where the CPU really excels. By [21:23] running both the control plane and the [21:25] inference on the CPU, you achieve [21:28] unified compute and eliminating the [21:31] latency of moving data to or from an [21:33] external accelerator. CPUs are also [21:36] built for reasoning and orchestration [21:38] where the high-speed single thread tasks [21:41] are needed to parse JSON, for example, [21:44] run Python code or execute tools. Plus, [21:49] they offer the massive memory capacity [21:51] needed to keep rapidly growing agent [21:53] states and the RAM. The bottom line is [21:57] to start CPU first to simplify your [22:01] stack on a unified footprint which [22:04] offers you a you know in excess of six [22:06] times lower token costs compared with [22:10] you know typical cloud GPU instances. [22:13] And we can still scale to GPUs when you [22:16] know you require heavy model sizes in [22:19] the application. [22:23] And to bring all of the elements we went [22:26] through together today, let's wrap it up [22:29] with a fun demo of live visual [22:31] understanding. This is a live camera [22:34] feed of a high school basketball game [22:36] analyzed as it plays by the media [22:38] framework running on a single Intel 6 [22:40] server. Two pipelines in parallel are [22:43] running one video, one audio on the same [22:45] CPU. The video examines every frame of [22:48] the 1080p at 30 frames per second to [22:52] recognize the shots at the basket and [22:55] adds each one as a highlight clip on the [22:58] card. The audio pipeline turns the live [23:01] commentary into the running transcript [23:03] beside it. The panel at the top right [23:06] shows that all of this runs several [23:09] times faster than real time or using [23:12] only a few% of that server resource in [23:15] real time. [23:17] Now, all of this is great, but you make [23:19] this a reality. We want you to try it [23:21] with your media, your application in [23:23] your environment. You can start with our [23:26] GitHub repos, give the Media AI [23:29] framework a test ride, and of course, [23:31] we'd love to hear from you. Help us [23:34] improve it further. [23:37] And to wrap up, here's a good set of [23:39] resources um to refer to. And thank you [23:42] very much for listening.