QCI Integrates Photonic Solvers With GPUs to Speed Generative AI Models

From Hype to Real Utility : Making Quantum Useful for AI Today With Quantum Computing Inc. — Ai4

Ai4

Enterprises can achieve immediate performance gains in AI diffusion and signal processing using room-temperature photonic hardware without waiting decades for fault-tolerant quantum computers.

AI-assisted summary of From Hype to Real Utility : Making Quantum Useful for AI Today With Quantum Computing Inc. — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 15, 2026, this talk features Yong Meng Sua, Chief Technology Officer at Quantum Computing Inc. (QCI). According to the description, the presentation demonstrates how quantum-inspired and photonic computing deliver practical utility for modern AI and optimization workflows today without relying on distant fault-tolerant quantum systems.

## TL;DR Rather than waiting decades for error-corrected universal quantum computers, Quantum Computing Inc. deploys room-temperature photonic hardware like the Dirac-3 entropic solver to accelerate discrete mathematical kernels in AI and optimization workloads. Operating as a complementary co-processor alongside GPUs, the technology improves generative diffusion modeling, enhances structured sequence recovery, and denoises weak lidar signals across real-world deployments with NASA and commercial data centers. To overcome data transfer latency bottlenecks and reduce energy consumption, QCI is advancing 3D heterogeneous integrated photonics that combine thin-film lithium niobate, silicon photonics, and CMOS electronics directly alongside classical compute.

## Atmosphere & style The presentation carries a pragmatic, grounded, and refreshing tone that intentionally cuts through the speculative hype dominating the quantum computing industry. Yong Meng Sua presents with a calm, engineering-focused clarity, pivoting away from distant theoretical promises toward tangible hardware racks, rigorous classical benchmarks, and immediate commercial viability. Listening to the talk feels like an honest industry reality check, replacing futuristic abstractions with concrete architectural roadmaps and measurable real-world deployments.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=4s) Rather than waiting 5 to 25 years for fault-tolerant quantum machines to replace classical compute, room-temperature quantum photonics can immediately accelerate discrete mathematical kernels like optimization and signal recovery. - [02:26](https://www.youtube.com/watch?v=d3FThrXyXhw&t=146s) The Dirac-3 entropic solver operates at room temperature in a 5U rack form factor by utilizing an open quantum optical feedback loop that harnesses vacuum fluctuations as computational fuel. - [02:26](https://www.youtube.com/watch?v=d3FThrXyXhw&t=146s) Photonic neuromorphic computing processes dynamic non-linear optical states to handle real-time edge inference, pattern recognition, and time-series forecasting. - [05:46](https://www.youtube.com/watch?v=d3FThrXyXhw&t=346s) Integrating the Dirac-3 as a Quantum Reverse Kernel inside standard diffusion models improves FID by 46% and Inception Score by 61% without requiring any retraining of pre-trained neural weights. - [05:46](https://www.youtube.com/watch?v=d3FThrXyXhw&t=346s) Native cubic and higher-order polynomial solving on photonic hardware enables modeling complex multi-variable interactions, achieving an 89% valid sequence recovery rate in discrete diffusion compared to 44% classically. - [09:55](https://www.youtube.com/watch?v=d3FThrXyXhw&t=595s) Photonic quantum-AI filtering has been validated across five NASA projects to successfully separate weak lidar return signals from strong solar background noise. - [09:55](https://www.youtube.com/watch?v=d3FThrXyXhw&t=595s) Photonic solvers are physically deployed in commercial data centers to tackle real-world optimization problems, including protein folding for drug discovery with Tri-I TDI. - [13:03](https://www.youtube.com/watch?v=d3FThrXyXhw&t=783s) Quantum systems are designed as hybrid co-processors that complement GPU-driven AI pipelines across data acquisition, denoising, and edge optimization rather than replacing classical computing. - [13:03](https://www.youtube.com/watch?v=d3FThrXyXhw&t=783s) QCI is developing a 3D heterogeneous material stack combining Thin-Film Lithium Niobate (TFLN), silicon photonics, and CMOS electronics to cut power draw and operational latency. - [18:06](https://www.youtube.com/watch?v=d3FThrXyXhw&t=1086s) To resolve communication latency between external software stacks and the photonic solver core, photonic chips are being designed for direct co-packaging alongside GPUs.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=4s) Moving from Quantum Hype to Real Utility Today

Following an introduction by Peggy and an inspiring opening story from the previous speaker, Nolan, the focus turns to the fundamental purpose of technology: whether artificial intelligence or quantum computing, human innovation must ultimately serve the greater human good. Moving from aspirational visions to the practical realities—what is candidly described as returning "to the darker side of the world"—the discussion shifts directly to cutting through industry hype to make quantum technology genuinely useful for AI workflows today. After presenting standard SEC forward-looking statements, the session addresses the persistent questions that follow anyone working in the quantum space: "When will a fault-tolerant quantum computer break RSA?" or "When will quantum computers finally compete with GPUs?"

These questions are fundamentally flawed because they tether quantum computing to a distant horizon requiring millions of error-corrected physical qubits deployed 5, 15, or 25 years into the future. That theoretical framing creates severe strategic deployment paralysis for enterprises. A much more practical question is asking what quantum photonics hardware can deliver right now. Instead of attempting to replace general-purpose classical compute, quantum photonics can specifically target discrete computational kernels—such as combinatorial optimization, temporal inference, and signal recovery—operating entirely at room temperature. Crucially, these systems can be evaluated and benchmarked in a straightforward, practical manner directly against standard CPUs and GPUs so enterprises can determine whether the investment yields tangible return.

- **Fault-Tolerant Horizon**: Long-term quantum roadmaps frequently focus on 5- to 25-year timelines requiring millions of error-corrected qubits. - **Immediate Value Proposition**: Room-temperature quantum photonics hardware targets specialized mathematical acceleration rather than universal compute replacement. - **Evaluation Criteria**: Benchmarked against classical hardware metrics to assess commercial cost-effectiveness and performance gains.

---

[02:26](https://www.youtube.com/watch?v=d3FThrXyXhw&t=146s) QCI's Core Hardware: Dirac-3 Entropic Solver and Neuromorphic Computing

Quantum Computing Inc. (QCI) delivers two primary hardware platforms available either on-premises or via cloud subscription: the Dirac-3 entropic quantum optimization machine (a photonic analog solver) and a photonic neuromorphic reservoir computer. The Dirac-3 machine takes an inverted approach to traditional gate-based quantum computing. Rather than isolating pristine qubits from environmental noise in cryogenic setups, entropic quantum computing operates as an open quantum system that harnesses decoherence and quantum vacuum fluctuations as computational "fuel." Mathematical problems are encoded into an optical feedback loop mixed with vacuum fluctuations, evolving iteratively toward the global minimum energy state representing the optimal solution.

```mermaid flowchart LR accTitle: Entropic Photonic Quantum Computing Loop subgraph Input["Problem Formulation"] A["Encoded Problem<br/>(QUBO / Analog / Multi-level)"] end subgraph PhotonicCore["Optical Feedback Core (Room Temperature)"] B["Optical Feedback Loop"] C["Quantum Vacuum<br/>Fluctuations"] B <--> C end subgraph Output["Solution Convergence"] D["Iterative Evolution<br/>(25–100 Cycles)"] E["Optimal Ground State /<br/>Statistical Distribution"] D --> E end A --> B B --> D ```

In parallel, QCI's photonic neuromorphic platform encodes data into dynamic photonic states, leveraging rich non-linear optical properties for real-time time-series prediction, edge inference, pattern recognition, and signal reconstruction. Packaged in a compact 5U rack-mount form factor, the Dirac-3 operates at room temperature with low power consumption. It functions as a drop-in acceleration solver for NP-hard problems, discrete and continuous optimization, Quadratic Unconstrained Binary Optimization (QUBO), integer programming, and analog multi-level encoding.

- **Dirac-3 Form Factor**: Packaged as a standard 5U rack-mount photonic server operating at room temperature with low operational power draw. - **Algorithmic Scope**: Solves discrete, continuous, QUBO, integer, and analog optimization problems via qubit and multi-level photonic states. - **Execution Profile**: Typically completes 25, 50, or 100 feedback iterations to establish a robust statistical distribution of optimal ground states. - **Neuromorphic Capabilities**: Processes dynamic non-linear photonic states for temporal sequence prediction and edge AI inference.

---

[05:46](https://www.youtube.com/watch?v=d3FThrXyXhw&t=346s) Quantum Reverse Kernels and Discrete Diffusion in AI Workflows

The utility of photonic hardware becomes evident when integrating the Dirac-3 solver directly into modern generative AI pipelines. In standard diffusion architectures, neural networks iteratively denoise data across 25 to 100 steps to synthesize images or video. By introducing a programmable Quantum Reverse Kernel (QRK) powered by the Dirac-3 as an internal sampling layer, the system evaluates all possible solution spaces in parallel during each generation step. This architecture requires zero retraining—using existing pre-trained neural network weights without needing external optimization solvers bolted on—delivering a 46% improvement in Fréchet Inception Distance (FID) to match ground truth distributions and a 61% boost in Inception Score for enhanced structural clarity.

```mermaid flowchart TD accTitle: Quantum-Enhanced AI Generative Pipeline subgraph ClassicalStack["Standard Diffusion Pipeline"] A["Pre-Trained Neural Weights<br/>(No Retraining Required)"] B["Iterative Denoising Loop<br/>(25–100 Steps)"] A --> B end subgraph QuantumLayer["Dirac-3 Photonic Sampler"] C["Parallel Solution Sampling<br/>(Quantum Reverse Kernel)"] D["Higher-Order / Cubic<br/>Correlation Analysis"] B <--> C B <--> D end subgraph Results["Generated Output"] E["Synthetic Images / Discrete Sequences<br/>(46% Better FID, 61% Inception Boost)"] B --> E end ```

Beyond continuous image generation, discrete diffusion architectures model sequential structured data such as DNA sequences. Traditional sequence generators factorize dependencies and struggle beyond simple pairwise relationships. Leveraging native cubic and higher-order polynomial solvers within the Dirac platform enables the system to capture multi-variable interactions naturally. When testing sequence generation, the quantum-enhanced model achieved an 89% valid ground-truth sequence recovery rate compared to just 44% for standard mathematical models. Furthermore, drawing 15 candidate sequences at a time consistently yields sequences that satisfy triadic motif rules, underscoring that identifying high-order correlations is essential for accurate synthetic data and image reconstruction.

- **Generative Improvements**: Yields a 46% better FID score and a 61% higher Inception Score without modifying or retraining underlying neural weights. - **Sequence Reconstruction**: Achieves an 89% valid sequence generation rate matching ground truth versus 44% for standard mathematical models. - **High-Order Interactions**: Native cubic terms solve complex triadic motif structures without mathematical factorization approximations.

---

[09:55](https://www.youtube.com/watch?v=d3FThrXyXhw&t=595s) Real-World Deployments: NASA Lidar Denoising and Data Center Integration

Moving past synthetic datasets and toy models, QCI's photonic solvers have been evaluated on real-world observational data across five collaborative projects with NASA. NASA had long encountered severe bottlenecks when using standard AI models to denoise weak lidar signals recorded against overwhelming solar background noise; over-filtering stripped away essential signal features, whereas under-filtering left excessive noise. Integrating the Dirac-3 solver alongside AI denoising algorithms enabled precise modeling of background noise characteristics, isolating and removing structured environmental noise without sacrificing the underlying lidar return signal.

```mermaid flowchart LR accTitle: NASA Lidar Denoising Architecture subgraph SignalInput["Raw Optical Acquisition"] A["Weak Lidar Return"] B["Strong Sunlight Noise"] A & B --> C["Noisy Composite Signal"] end subgraph Processing["Hybrid Processing"] D["Classical AI Filter"] E["Dirac-3 Correlation Solver"] C --> D D <--> E end subgraph CleanOutput["Reconstructed Signal"] F["Denoised Lidar Profile<br/>(Preserved Critical Features)"] E --> F end ```

Transitioning quantum systems from laboratory demonstrations to customer environments remains a critical hurdle across all quantum modalities. QCI has demonstrated physical installations of the Dirac-3 system inside operational commercial data centers, including the Quantum Corridor Quantum-Safe Network in Indiana and The Forge at the Center for Federal Services Research in Washington, D.C. Within these data center rack environments, collaborators such as the Tri-Institutional Therapeutics Discovery Institute (Tri-I TDI) leverage the platform to solve complex protein folding optimization problems, assisting molecular structural discovery for targeted drug development.

- **NASA Lidar Denoising**: Successfully extracts weak return signals submerged in high-intensity solar background noise across five NASA initiatives. - **Data Center Deployments**: Operational rack-mounted installations at Indiana's Quantum Corridor and Washington, D.C.'s The Forge. - **Biomedical Application**: Utilized by the Tri-Institutional Therapeutics Discovery Institute for conformational optimization in protein folding and drug discovery.

---

[13:03](https://www.youtube.com/watch?v=d3FThrXyXhw&t=783s) The Hybrid Quantum-AI Stack and Integrated Photonics Architecture

Quantum computing is not designed to replace AI or classical GPUs; it is built to operate as a complementary co-processor across hybrid pipelines. In this architecture, classical sensors capture raw inputs, a quantum-AI solver denoises the signal, GPUs handle core neural network training and inference, and a quantum layer manages complex optimization and decision-making at the edge. To separate commercial reality from hype, enterprises must evaluate five pragmatic criteria:

| Evaluation Criterion | Practical Requirement | |:--- |:--- | | **Deployment Timeline** | Can the solution run in production within the current fiscal quarter? | | **Classical Benchmarking** | Does it measurably outperform classical CPU/GPU baselines on identical workloads? | | **Problem Relevance** | Does it address an immediate, high-value business bottleneck? | | **Infrastructure Integration** | Can it integrate cleanly into existing IT and data center environments? | | **Energy Economics** | Does the performance gain justify the operational power and cost footprint? |

```mermaid flowchart TD accTitle: 3D Integrated Photonics Stack subgraph Stack["3D Heterogeneous Material Stack"] L1["Thin-Film Lithium Niobate (TFLN)<br/>(Unmatched Non-Linear Optics & Electro-Optics)"] L2["Silicon Photonics Layer<br/>(Scalable Waveguides & Optical Interconnects)"] L3["CMOS Electronics Base<br/>(Standard High-Speed Logic & Control)"] L1 <--> L2 L2 <--> L3 end ```

To eliminate energy and latency overheads, QCI is pursuing 3D heterogeneous photonic integration. Following the acquisition of a specialized semiconductor facility, the company is combining thin-film lithium niobate (TFLN)—leveraging 15 years of material development for non-linear optical properties—with silicon photonics and standard CMOS electronics. This domestic U.S. supply chain enables high-density electro-optical interconnects that dramatically reduce energy per bit. Organizations can validate these capabilities quickly through structured 90-day pilot programs, testing specific computational kernels against classical baselines before scaling up deployment.

- **Core Takeaway 1**: Real-world quantum utility does not require waiting decades for fault-tolerant universal machines. - **Core Takeaway 2**: Quantum architectures naturally complement and accelerate GPU-driven AI pipelines rather than competing with them. - **Core Takeaway 3**: Performance must be judged honestly on immediate benchmark improvements rather than multi-decade promises. - **Pilot Implementation**: 90-day structured proof-of-concept projects allow organizations to identify viable kernels and measure concrete ROI.

---

[18:06](https://www.youtube.com/watch?v=d3FThrXyXhw&t=1086s) Q&A: Overcoming Latency with Co-Packaged Photonic Chips

In the concluding question-and-answer exchange, an attendee asks about the real-world cost-effectiveness and practical complexity of integrating these photonic systems into industrial environments. The response highlights the primary operational bottleneck currently limiting hybrid systems: data transfer latency. While the internal Dirac solver demonstrates clear algorithmic improvements over classical routines, streaming data back and forth between client software, embedded control layers, and the physical photonic core introduces communication overhead where "time and money are wasted."

To eliminate these latency barriers and make scaling cost-effective, QCI is advancing its 3D integrated photonic chip platform. By packaging thin-film lithium niobate and silicon photonics directly alongside GPUs using optical interconnect architectures similar to modern data center networking, the system avoids external software-stack delays. Co-packaging the photonic hardware directly with classical compute processors removes intermediate communication bottlenecks, paving the way for low-latency, cost-efficient quantum acceleration in production AI workloads.

- **Latency Bottleneck**: Current hybrid setups incur latency transferring data between client software, embedded layers, and the optical solver. - **Co-Packaging Solution**: 3D integrated photonic chips designed to sit directly adjacent to GPUs to eliminate interconnect delays. - **Industrial Scaling**: Direct optical co-packaging reduces transmission overhead, bringing down the total operational cost of quantum-accelerated AI.

## Notable quotes > Quantum computing is often positioned as a future breakthrough, but meaningful value is already being realized today. — Yong Meng Sua

> These technologies are not replacements for classical AI, but powerful complements, particularly for complex optimization, pattern recognition, and data-intensive decision-making. — Yong Meng Sua

> Join this session to separate signal from noise and start leveraging quantum for practical impact today. — Yong Meng Sua

> [00:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=4s) to the darker side of the world — Yong Meng Sua

> [00:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=4s) When will a fault-tolerant quantum computer break RSA? — Yong Meng Sua

> [00:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=4s) When will quantum computers finally compete with GPUs? — Yong Meng Sua

> [02:26](https://www.youtube.com/watch?v=d3FThrXyXhw&t=146s) fuel — Yong Meng Sua

> [18:06](https://www.youtube.com/watch?v=d3FThrXyXhw&t=1086s) time and money are wasted — Yong Meng Sua

## Mentioned Yong Meng Sua, Peggy, Nolan, Quantum Computing Inc. (QCI), Ai4, Dirac-3, Quantum Reverse Kernel (QRK), NASA, Quantum Corridor, The Forge, Center for Federal Services Research, Tri-Institutional Therapeutics Discovery Institute (Tri-I TDI), Thin-Film Lithium Niobate (TFLN), CMOS, GPU, CPU, RSA

## Video description Yong Meng Sua - Chief Technology Officer, Quantum Computing Inc.

Quantum computing is often positioned as a future breakthrough, but meaningful value is already being realized today.

In this talk, I will highlight how quantum-inspired and photonic computing approaches are delivering practical solutions to real-world AI and optimization challenges, without relying on fault-tolerant quantum systems. These technologies are not replacements for classical AI, but powerful complements, particularly for complex optimization, pattern recognition, and data-intensive decision-making. Drawing on real world deployment, I will discuss where quantum and photonics approaches are already gaining traction, how they complement existing AI workflows, and what differentiates practical solutions from theoretical promise. Join this session to separate signal from noise and start leveraging quantum for practical impact today.

[Source](https://www.youtube.com/watch?v=d3FThrXyXhw)

中文简报

QCI将光子求解器与GPU集成以加速生成式AI模型 企业无需等待数十年的容错量子计算机问世，即可利用室温光子硬件在AI扩散和信号处理中实现即时性能提升。 Quantum Computing Inc. 无需等待数十年的纠错通用量子计算机，而是部署了如 Dirac-3 熵求解器等室温光子硬件，以加速 AI 和优化工作负载中的离散数学内核。作为与 GPU 协同运行的互补协处理器，该技术在与 NASA 及商业数据中心的实际部署中，改善了生成式扩散建模，增强了结构化序列恢复，并对微弱的激光雷达信号进行了降噪。为了克服数据传输延迟瓶颈并降低能耗，QCI 正在推进 3D 异构集成光子技术，将薄膜铌酸锂、硅光子以及 CMOS 电子元件直接与经典计算集成在一起。 与其等待5到25年后的容错量子计算机，常温量子光子硬件已能即刻加速离散数学核心任务；QCI的Dirac-3熵增求解器采用5U机架形态，通过开放式光学反馈回路并利用量子真空涨落作为算力燃料。 (0:04) 将Dirac-3作为量子逆内核（Quantum Reverse Kernel）集成至标准扩散模型中，无需重新训练既有神经网络权重，即可将FID指标优化46%，并将Inception Score提升61%。 (5:46) 光子硬件支持原生三次及高阶多项式求解，在离散扩散序列生成中实现了89%的有效序列恢复率，远超经典数学模型的44%。 (5:46) 光子求解器已在五个NASA项目中得到验证，成功从强烈太阳背景噪声中提取微弱的激光雷达回波，并在商业数据中心部署运行，协助Tri-I TDI开展药物发现中的蛋白质折叠优化。 (9:55) 量子系统定位为与GPU协同工作的混合协处理器，QCI正在推进基于薄膜铌酸锂（TFLN）的3D异构集成光子芯片，通过光互连直接与经典芯片共封装来消除数据传输延迟瓶颈。 (13:03)

Reading summary

Quantum Computing Inc. argues that commercial quantum utility does not require waiting decades for fault-tolerant machines, advocating instead for room-temperature photonic co-processors that accelerate specialized mathematical bottlenecks in current AI workflows. Targeting Immediate Acceleration over Distant Horizons (0:04) Long-term quantum roadmaps frequently project timelines of 5 to 25 years requiring millions of error-corrected physical qubits to achieve universal fault tolerance. QCI asserts that tethering quantum strategy to these distant horizons causes deployment paralysis for enterprises. Rather than attempting to replace general-purpose classical compute, the company proposes using room-temperature quantum photonics to target discrete mathematical kernels, such as combinatorial optimization, temporal inference, and signal recovery. Under this framework, photonic hardware is treated as an immediate domain-specific accelerator. Instead of relying on speculative future performance, systems are evaluated through direct benchmarking against standard CPUs and GPUs to measure whether commercial cost-effectiveness and performance gains justify integration. Open-System Entropic and Neuromorphic Computing (2:26) QCI's hardware portfolio centers on the Dirac-3 entropic quantum solver and a photonic neuromorphic platform. Unlike gate-based cryogenic processors that isolate qubits from environmental disturbances, the Dirac-3 functions as an open quantum system that uses decoherence and quantum vacuum fluctuations as computational fuel. Mathematical formulations—including Quadratic Unconstrained Binary Optimization (QUBO), integer programming, and continuous analog problems—are encoded into an optical feedback loop, evolving across 25 to 100 cycles to establish a statistical distribution of optimal ground states. The Dirac-3 is packaged as a room-temperature 5U rack-mount unit designed for low power consumption. Running in parallel, QCI's neuromorphic reservoir computing system harnesses dynamic non-linear optical states to perform real-time time-series forecasting, edge inference, and signal reconstruction without cryogenic cooling. Enhancing Generative AI and High-Order Modeling (5:46) In generative AI architectures, the Dirac-3 integrates into diffusion pipelines as a programmable Quantum Reverse Kernel (QRK). By acting as an internal sampling layer that evaluates solution spaces in parallel across 25 to 100 denoising iterations, the system utilizes pre-trained neural network weights without requiring model retraining. QCI reports that this hybrid approach yields a 46% improvement in Fréchet Inception Distance (FID) and a 61% increase in Inception Score relative to standard diffusion baselines. For structured discrete data, such as DNA sequences, conventional models factorize dependencies and struggle with multi-variable interactions. Dirac-3 provides native polynomial solvers for cubic and higher-order terms, enabling it to satisfy triadic motif rules. In sequence recovery tests drawing 15 candidates at a time, the quantum-assisted model achieved an 89% valid ground-truth recovery rate compared to 44% for standard mathematical methods. Field Deployments in Denoising and Structural Discovery (9:55) QCI has tested its photonic systems on production data and data center environments. Across five collaborative initiatives with NASA, the Dirac-3 was deployed alongside AI denoising algorithms to recover weak lidar signals obscured by strong solar background noise. The solver modeled structured background interference to isolate environmental noise without removing underlying signal features, resolving an over- and under-filtering trade-off encountered by conventional AI filters. The hardware has also been deployed in commercial and research data centers, including Indiana's Quantum Corridor and The Forge in Washington, D.C. Within these facilities, partners such as the Tri-Institutional Therapeutics Discovery Institute (Tri-I TDI) utilize the platform's optimization capabilities for molecular conformation and protein folding to support targeted drug design. Mitigating Latency Through 3D Photonic Co-Packaging (13:03) In practice, hybrid quantum-classical architectures encounter latency bottlenecks caused by transferring data between client software, embedded control layers, and the photonic core. To reduce transmission overhead, QCI is pursuing 3D heterogeneous photonic integration. This platform combines thin-film lithium niobate (TFLN)—utilized for its non-linear optical properties—with silicon photonics and CMOS logic manufactured through a domestic U.S. facility. The resulting architecture is designed to co-package photonic processing chips directly alongside classical GPUs using high-density optical interconnects, bypassing external software-stack delays. To validate operational viability and energy economics, QCI recommends that organizations run structured 90-day pilot programs targeting specific computational kernels before scaling deployment. While QCI demonstrates near-term performance gains across denoising, sequence modeling, and optimization, the long-term scalability of its hybrid model depends on eliminating data transfer bottlenecks by successfully co-packaging photonic chips directly with classical GPUs.

精炼阅读摘要

Quantum Computing Inc.（QCI）主张，量子计算在商业落地中无须等待耗时数十年且依赖数百万容错物理量子比特的通用机器，而应利用室温光量子硬件作为辅助协处理器，在当下针对特定数学核心加速人工智能工作流。 理念转向：以室温光量子加速特定算子 (0:04) 长期以来，量子技术路线图往往锚定在5至25年之后实现拥有数百万个纠错物理量子比特的容错量子计算机上。演讲者指出，这种将量子计算绑定于远期宏大目标的理论视角，容易导致企业在实际战略部署中陷入停滞。与其纠结于量子计算机何时能破解RSA加密算法或何时全面取代GPU，更具商业价值的切入点在于审视现有光量子硬件能够立即输出哪些算力支持。 该方案并不寻求全盘替代通用经典算力，而是将室温光量子硬件专门用于组合优化、时序推理和信号恢复等特定离散计算核心。光量子系统的有效性可直接对照标准CPU与GPU的运行指标进行性能基准测试与成本效益核算，帮助企业评估投资在当下是否能换取可量化的实际回报。 核心硬件：Dirac-3熵求解器与神经形态 (2:26) QCI推出的核心硬件包括Dirac-3熵量子优化求解器与光子神经形态储层计算平台，支持本地机房部署或云端订阅接入。不同于依靠极低温环境隔离环境噪声的传统门保真量子计算方案，Dirac-3采用开放量子系统架构，反向利用退相干与量子真空涨落作为计算驱动力。数学问题被编码至与真空涨落耦合的光学反馈环路中，系统经过25至100次迭代演化即可收敛至代表最优解的基态能量分布。 硬件系统封装为标准5U机架式服务器，具备室温运行与低运行功耗的特征，可作为即插即用求解器处理NP-hard难题、离散与连续优化、二次无约束二值优化（QUBO）、整数规划以及模拟多能级编码。与其并行的光子神经形态系统则利用非线性光学动态特性，处理实时时间序列预测与边缘AI推理任务。 赋能生成式AI：量子反向核与离散扩散 (5:46) 在生成式AI工作流中，Dirac-3求解器可作为可编程量子反向核（QRK）直接嵌入扩散模型的生成循环。该架构无须重新训练模型，即可直接沿用现有的预训练神经网络权重，在去噪步骤中并行评估全解空间。测试数据表明，该方案使衡量图像生成质量的Fréchet Inception Distance（FID）指标改善了46%，并将Inception Score提升了61%。 在建模DNA序列等离散结构化数据时，传统序列生成模型因依赖数学因式分解近似而难以处理高阶多变量依赖。Dirac-3依托其原生的三次及更高阶多项式求解能力直接刻画复杂关联。在序列生成测试中，量子增强模型实现了89%的真实序列恢复率，显著高于传统数学模型的44%，且在一次抽取15条候选序列时能稳定符合三元基序规则。 真实落地部署：NASA降噪与数据中心集成 (9:55) 在实际工程验证方面，QCI在与NASA合作的5个项目中测试了光量子求解器的有效性。针对强日光背景噪声下微弱激光雷达（Lidar）信号极难提取的瓶颈，联合算法利用Dirac-3精确建模环境背景噪声特征，在保留雷达回波关键特征的同时滤除了结构化强噪声，避免了传统AI滤波容易出现的过滤波或欠滤波失真。 在物理设施集成上，Dirac-3已进入商用数据中心机架，包括印第安纳州的Quantum Corridor抗量子网络节点以及华盛顿特区联邦服务研究中心的The Forge设施。在此类算力环境中，三机构治疗药物研发所（Tri-I TDI）正借助该平台解决复杂的蛋白质折叠构象优化问题，以此辅助分子结构分析与靶向药物开发。 架构融合演进：3D光电共封装化解传输延迟 (13:03) 在QCI规划的混合量子-AI流水线中，传感器负责原始数据采集，量子AI求解器执行信号降噪，GPU专注核心神经网络训练与推理，量子层则在边缘端承担复杂组合优化与决策。企业在落地评估时应遵循五个务实维度：当前财季能否投产、相对经典硬件是否有可测量的性能优势、是否对应高价值业务瓶颈、能否平滑融入IT基础设施，以及能耗与成本是否具备经济性。企业通常可通过结构化的90天试点项目验证具体算子。 针对混合系统在客户端软件、控制层与光子内核之间数据往返传输产生的通信延迟瓶颈，QCI正在推进3D异构光电集成技术。通过整合具备非线性光学特性的薄膜铌酸锂（TFLN）、硅光子波导及标准CMOS电子层，QCI计划将光子芯片直接与GPU共封装（Co-Packaging），通过光互连规避外部软件栈开销，降低系统延迟与单位比特能耗。 QCI展示了室温光量子硬件在图像生成、序列重构与信号降噪等离散AI算子上的加速能力，但该路线目前定位于特定数学计算的辅助加速而非通用量子计算；系统能否在工业规模实现广泛落地，仍取决于光子芯片与GPU共封装技术的推进速度，以及能否在工程上彻底解决异构计算架构之间的数据传输延迟瓶颈。

Detailed notes — zh

## Background 该视频由行业会议频道 Ai4 于 2026 年 8 月 15 日在 YouTube 平台发布。内容主讲人为 Quantum Computing Inc.（QCI）首席技术官 Yong Meng Sua，视频围绕量子与光子计算如何摆脱对遥远容错系统的依赖，在当下直接赋能现实 AI 优化、模式识别与决策工作流展开探讨。 ## TL;DR 演讲倡导跳出数十载后才能实现的“容错量子霸权”炒作，转而聚焦于能在室温、低功耗机架环境下稳定运行的光子模拟与熵量子硬件（如 Dirac 系列）。通过将光学反馈环路与开放耗散机制转化为算力，QCI 展示了其在量子反向核（QRK）图像扩散去噪与高阶离散序列生成中的显著增益，FID 改善达 46%。结合 NASA 激光雷达去噪及数据中心机架落地的真实案例，Sua 提出了评判量子项目落地价值的五大准则，并展望了通过薄膜铌酸锂（TFLN）3D 光电单片集成与 GPU 共封装（CPO）来彻底攻克数据传输延迟的演进路线。 ## Atmosphere & style 整场演讲呈现出一种极为务实克制且直击痛点的技术现实主义基调，讲者完全摒弃了空洞的宏大叙事，以清醒冷静的姿态直斥行业内的虚假繁荣与长期画饼。叙述节奏紧凑严密，从硬核的光学物理演化机理层层递进到 AI 采样指标、实战部署数据以及芯片制造工艺，工程细节极其扎实。听感上专业而笃定，既坦诚指出当前板级互连延迟等物理瓶颈，又用清晰的 ROI 对标与实测数据为量子计算融入现代 AI 基础设施指明了立即可行的工程路径。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=4s) 行业应摆脱对数百万纠错量子比特和遥远容错时代的虚假炒作，转而关注量子技术在当下能为优化、时序推理等特定计算核心带来的实际加速与商业投资回报率（ROI）。 - [02:26](https://www.youtube.com/watch?v=d3FThrXyXhw&t=146s) 熵量子计算（Entropic Quantum Computing）颠覆了传统门模型对环境完全隔离的思路，反向利用开放系统的耗散与退相干作为演化动力，驱动光学反馈环路自发收敛至 NP-hard 问题的基态最优解。 - [02:26](https://www.youtube.com/watch?v=d3FThrXyXhw&t=146s) Dirac 系列光子求解器采用标准的 5U 机架式服务器形态，完全支持室温低功耗运行，可作为即插即用（Drop-in）组件嵌入现有 AI 技术栈进行调度与资源决策。 - [05:46](https://www.youtube.com/watch?v=d3FThrXyXhw&t=346s) 在连续扩散图像生成中，量子反向核（QRK）作为并行采样层直接接入预训练模型，无需重新训练权重即可使 FID 分数改善 46%，Inception 分数提升约 61%。 - [05:46](https://www.youtube.com/watch?v=d3FThrXyXhw&t=346s) 在离散序列生成任务中，Dirac 的原生高阶项求解能力可有效捕捉高阶关联，将符合真实基准的有效序列生成率从传统数学模型的 44% 显著提升至 89%。 - [09:55](https://www.youtube.com/watch?v=d3FThrXyXhw&t=595s) 与 NASA 合作的 LiDAR 降噪项目证明，将量子光子层融入 AI 降噪流水线，能在强日光干扰下精确剥离背景噪声而不损伤极微弱的返回信号几何特征。 - [09:55](https://www.youtube.com/watch?v=d3FThrXyXhw&t=595s) QCI 硬件已成功走出实验室，直接部署于商用数据中心机架（如 Quantum Corridor 与 The Forge），并在生物医药领域的蛋白质折叠结构优化中投入实际应用。 - [09:55](https://www.youtube.com/watch?v=d3FThrXyXhw&t=595s) 量子计算的核心定位是赋能并补充经典 AI 计算流水线，而非取代 CPU/GPU：由经典硬件负责感知与模型训练，量子层专注前端去噪与后端决策优化。 - [14:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=844s) 评估量子技术落地价值需遵循五大务实标准：季度内执行周期、直接硬碰经典 CPU/GPU 集群的对标、解决真实商业痛点、低摩擦 IT 集成以及随规模扩展的能源经济性。 - [14:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=844s) 通过 3D 混合单片集成技术，将薄膜铌酸锂（TFLN）的超高速强非线性调制与成熟硅光子制造生态及底层 CMOS 电控结合，构建起高性能且自主可控的光子计算供应链。 - [18:06](https://www.youtube.com/watch?v=d3FThrXyXhw&t=1086s) 当前外置光子硬件的主要瓶颈在于跨网络与总线的数据搬运延迟与通信开销，演进方向是推进光子核心与 GPU 的共封装（CPO）以彻底消除板级通信延迟。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=4s) 破除量子霸权炒作：关注当下的实用价值 Quantum Computing Inc.（QCI）首席技术官 Yong Meng Sua 在开场时首先感谢了主持人 Peggy 和上一位演讲者 Nolan。他赞同人类创造的技术（无论是人工智能还是量子计算）都应当服务于人类的更大福祉。然而，他坦言自己必须把大家带回“更现实的层面”，探讨如何摆脱浮躁的炒作，转向探讨量子计算在今天如何真正为 AI 提供实用价值（utility）。 行业中人们最常问的问题往往是：“容错量子计算机何时能够破解现有的加密体系（如 RSA）？”或者“量子计算机何时能在算力上全面击败 GPU？”Sua 指出，这完全是问错了方向。这种思维总是将目光锚定在需要数百万个纠错量子比特、需要等待 5 年、15 年甚至 25 年的遥远未来，这让企业很难在当下制定落地策略。更切实际的问题应当是：“量子技术在今天能够提供什么？”通过 QCI 所倡导的量子光子硬件，人们无需等待容错系统的成熟，就能针对优化、时序推理（temporal inferences）或信号恢复等特定计算核心（kernel）获得实际加速，且所有设备均可在室温下稳定运行。 - 行业误区：过分执着于数百万纠错量子比特与遥远的容错时代，导致当下战略部署困难。 - 现实路径：利用量子光子硬件，直接针对优化、时序推理和信号恢复等特定算子，在室温环境下实现即时可用性。 - 评估标准：用“最枯燥但客观”的方式与经典 CPU、GPU 直接对标，衡量投资回报率（ROI）。 --- ### [02:26](https://www.youtube.com/watch?v=d3FThrXyXhw&t=146s) 熵量子计算与 Dirac 系列光子模拟求解器 QCI 目前为市场提供两种核心硬件方案，均支持本地部署（On-Premise）或云端订阅访问。第一种是熵量子优化机（Entropic Quantum Computing），即 **Dirac** 系列光子模拟求解器；第二种是用于神经形态推理的光子世界计算机（Photonic World Computer），它将信息编码进动态光子态中，利用其丰富的非线性特征进行实时时间序列预测和信号级 AI 重建。 ```mermaid flowchart TD accTitle: 熵量子计算与光子反馈环路 subgraph Input["输入阶段"] A["业务优化问题 (QUBO / 整数 / 模拟)"] B["量子真空涨落 (Quantum Vacuum Fluctuation)"] end subgraph Loop["光子演化核心"] C["光学反馈环路 (Optical Feedback Loop)"] D["开放量子系统耗散与退相干调控"] end subgraph Output["结果输出"] E["基态收敛 (最优解分布)"] F["5U 机架式光子服务器 (室温低功耗运行)"] end A --> C B --> C C <--> D D --> E E --> F ``` Sua 重点拆解了熵量子计算的工作原理。传统门模型（Gate-based）量子计算极力将光子量子比特与外界环境完全隔离以避免退相干，但熵量子计算反其道而行之：它将系统设计为开放量子系统，把退相干和耗散作为动力学演化的“燃料”而非阻碍。系统将问题编码后引入光学反馈环路中，与量子真空涨落混合，经过 25、50 或 100 轮快速迭代演化，驱动系统自发耗散并收敛至基态，从而直接输出 NP-hard 问题的全局最优解分布。 - **Dirac 3 硬件形态**：紧凑的 5U 机架式光子服务器，完全在室温下运行，具备极低功耗。 - **问题覆盖面**：支持离散和连续优化、QUBO、整数规划求解以及基于连续变量（多能级编码）的光子模拟求解。 - **软件集成度**：可作为即插即用（Drop-in）求解器嵌入现有 AI 技术栈，直接用于调度、资源分配、投资组合优化、传感器布局和推理层决策。 --- ### [05:46](https://www.youtube.com/watch?v=d3FThrXyXhw&t=346s) 量子增强扩散模型：连续采样与离散序列生成 在结合现代 AI 架构方面，Sua 展示了两个前沿模型应用。第一个是基于 **量子反向核（Quantum Reverse Kernel, QRK）** 的扩散采样模型。标准的连续扩散模型在生成图像或视频时，需要神经网络进行数十次迭代去噪逐步逼近解。QRK 将 Dirac 作为采样层（Sampler Layer）直接嵌入迭代环路中，同时对所有潜在解空间进行并行采样。 ```mermaid flowchart LR accTitle: 量子反向核与离散扩散架构对比 subgraph QRK["连续图像生成: QRK 架构"] direction TB Q1["预训练神经网络权重 (无需重新训练)"] --> Q2["Dirac 并行采样层"] Q2 --> Q3["高质量图像重建 (FID 降低 46%)"] end subgraph DD["离散序列生成: 高阶关联求解"] direction TB D1["离散数据序列 (如 DNA / 像素编码)"] --> D2["Dirac 原生三次项求解器 (Native Cubic Solver)"] D2 --> D3["捕捉高阶关联 (满足三元基序规则)"] end ``` 这种采样方式无需重新训练基础神经网络权重，也无需在外围挂载复杂的优化器。测试结果显示，QRK 架构的 FID（Fréchet Inception Distance）得分改善了 46%（更接近真实分布），Inception 分数提升了约 61%，生成的图像细节和清晰度均大幅提升。 第二个应用是 **离散扩散（Discrete Diffusion）** 模型，用于生成类似 DNA 序列的离散数据序列。传统序列模型通常采用因式分解来近似计算，只能处理成对（pairwise）关联，而 Dirac 的原生高能级求解能力（Cubic / 高阶相互作用项）能够直接捕捉序列中的高阶多体关联： - **有效序列生成率**：反映真实基准的有效序列生成比例从标准数学模型的 44% 显著提升至 89%。 - **采样效率**：每次仅需抽取 15 个候选序列，即可确保其中至少有一个完美符合真实分布。 - **三元基序规则（Triadic Motif Rule）**：激活硬件原生高阶项后，系统能够完全满足复杂的三元约束规则。 --- ### [09:55](https://www.youtube.com/watch?v=d3FThrXyXhw&t=595s) 真实场景落地：NASA 激光雷达去噪与数据中心部署 从实验室原型走向实际业务场景，Sua 分享了 QCI 与 NASA 合作开展的项目经验。NASA 在使用传统 AI 进行激光雷达（LiDAR）信号去噪时面临严峻的权衡困境：“去噪过度会丢失关键几何特征，去噪不足则信噪比极差”。在强日光背景干扰下，微弱的激光雷达返回信号几乎被噪声淹没。通过将 Dirac 作为量子层融入 AI 降噪流水线中，算法能够精准建模背景光噪声的统计关联，在不损伤真实微弱信号的前提下实现了高精度的背景噪声剥离。 | 落地维度 | 实施细节与合作案例 | |:--- |:--- | | **基础设施部署** | 成功将 Dirac 5U 硬件设备直接安装进标准商用数据中心机架，包括印第安纳州的 Quantum Corridor 量子安全网络，以及华盛顿特区的 The Forge（中央联邦服务研究中心）。 | | **生物医药应用** | 与三机构治疗发现研究所（Tri-Institutional TDI）合作，利用部署在机架上的量子光子模拟求解器进行复杂蛋白质折叠结构优化，加速药物研发。 | | **混合架构协作** | 坚持“量子赋能而非取代经典”原则：经典传感器输入 $\to$ 量子+AI 协同去噪 $\to$ GPU 模型训练与常规推理 $\to$ 量子层实施最终决策优化与边缘控制。 | --- ### [14:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=844s) 鉴别炒作的五大标准与薄膜铌酸锂 3D 光电集成 为了帮助行业在技术选型中“去伪存真”，Sua 提出了评估量子项目的五大核心准则： 1. **执行周期**：能否在本季度内实际运行并纳入业务预算？ 2. **基准对标**：能否与最顶尖的经典 CPU/GPU 集群进行硬碰硬的公开对标？ 3. **真实痛点**：是否真正解决了能创造商业价值的核心问题？ 4. **系统集成**：能否低摩擦融入现有的 IT 和 AI 基础设施栈？ 5. **能源经济性**：单位比特的算力和能耗成本是否支持业务随规模扩展？ ```mermaid flowchart TD accTitle: 3D光电单片集成栈 subgraph Stack["3D 薄膜铌酸锂光电集成栈"] L1["薄膜铌酸锂层 (TFLN): 卓越的电光效应与非线性光学物理"] L2["硅光子层 (Silicon Photonics): 依托成熟半导体制造生态"] L3["底层 CMOS 电子控制层: 信号互联与逻辑驱动"] end L1 <==>|光学/电学校准互连| L2 L2 <==>|硅通孔 / 高速互连| L3 ``` 为解决光子计算的扩展瓶颈，QCI 通过近期对一家半导体公司的战略并购，全力推进 **薄膜铌酸锂（Thin-Film Lithium Niobate, TFLN）** 与硅光子及 CMOS 电子层的 3D 混合单片集成。单靠 TFLN 缺乏现有硅光庞大的制造生态，而单靠硅光则无法提供超高速的电光调制与强非线性响应。二者的 3D 垂直整合不仅能将能效比推进到极致，还构建了完整的美国本土自主供应链。Sua 建议企业无需盲目投入数十亿美元，可以通过 90 天的试点项目（POC），从小切口算子开始严格对标 AI+量子的实际增益。 --- ### [18:06](https://www.youtube.com/watch?v=d3FThrXyXhw&t=1086s) 现场 Q&A：解决数据搬运延迟与 GPU 共封装演进 在问答环节，现场观众针对硬件架构的复杂性与当前工业落地的成本效益提出了疑问。 Sua 坦率指出了现有外置系统的物理瓶颈：虽然在 AI 模型中加入 Dirac 求解器能带来显著的算法精度提升，但在目前通过网络或总线进行数据流传输时，数据从客户端软件经由嵌入式固件传输至光子核心并返回，会产生较大的转换延迟与通信开销，这在规模化场景下会推高成本。 这也是 QCI 大力押注 3D 集成光子芯片的根本原因。未来的演进方向是采用与数据中心光互连相似的架构，实现光子硬件核心与 GPU 的 **共封装（Co-Packaged Optics, CPO）**，直接消除板级通信延迟，从而在超大规模 AI 推理与决策层中实现极高的成本效益。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=4s) 容错量子计算机何时能够破解现有的加密体系（如 RSA）？ — Yong Meng Sua > [00:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=4s) 量子计算机何时能在算力上全面击败 GPU？ — Yong Meng Sua > [00:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=4s) 量子技术在今天能够提供什么？ — Yong Meng Sua > [00:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=4s) 用‘最枯燥但客观’的方式与经典 CPU、GPU 直接对标，衡量投资回报率（ROI）。 — Yong Meng Sua > [02:26](https://www.youtube.com/watch?v=d3FThrXyXhw&t=146s) 它将系统设计为开放量子系统，把退相干和耗散作为动力学演化的‘燃料’而非阻碍。 — Yong Meng Sua > [09:55](https://www.youtube.com/watch?v=d3FThrXyXhw&t=595s) 去噪过度会丢失关键几何特征，去噪不足则信噪比极差 — Yong Meng Sua > [09:55](https://www.youtube.com/watch?v=d3FThrXyXhw&t=595s) 量子赋能而非取代经典 — Yong Meng Sua > [14:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=844s) 能否在本季度内实际运行并纳入业务预算？ — Yong Meng Sua > [14:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=844s) 能否与最顶尖的经典 CPU/GPU 集群进行硬碰硬的公开对标？ — Yong Meng Sua > [14:04](https://www.youtube.com/watch?v=d3FThrXyXhw&t=844s) 企业无需盲目投入数十亿美元，可以通过 90 天的试点项目（POC），从小切口算子开始严格对标 AI+量子的实际增益。 — Yong Meng Sua ## Mentioned Quantum Computing Inc. (QCI), Yong Meng Sua, Peggy, Nolan, Dirac / Dirac 3, 熵量子计算 (Entropic Quantum Computing), 光子世界计算机 (Photonic World Computer), 量子反向核 (Quantum Reverse Kernel, QRK), 离散扩散 (Discrete Diffusion), NASA, Quantum Corridor, The Forge, 三机构治疗发现研究所 (Tri-Institutional TDI), 薄膜铌酸锂 (Thin-Film Lithium Niobate, TFLN), 共封装光学 (Co-Packaged Optics, CPO) ## Video description Yong Meng Sua - Chief Technology Officer, Quantum Computing Inc. Quantum computing is often positioned as a future breakthrough, but meaningful value is already being realized today. In this talk, I will highlight how quantum-inspired and photonic computing approaches are delivering practical solutions to real-world AI and optimization challenges, without relying on fault-tolerant quantum systems. These technologies are not replacements for classical AI, but powerful complements, particularly for complex optimization, pattern recognition, and data-intensive decision-making. Drawing on real world deployment, I will discuss where quantum and photonics approaches are already gaining traction, how they complement existing AI workflows, and what differentiates practical solutions from theoretical promise. Join this session to separate signal from noise and start leveraging quantum for practical impact today. [Source](https://www.youtube.com/watch?v=d3FThrXyXhw)

Source quotes

[0:04] Yong Meng Sua: 容错量子计算机何时能够破解现有的加密体系（如 RSA）？ [0:04] Yong Meng Sua: 量子计算机何时能在算力上全面击败 GPU？ [0:04] Yong Meng Sua: 量子技术在今天能够提供什么？ [0:04] Yong Meng Sua: 用‘最枯燥但客观’的方式与经典 CPU、GPU 直接对标，衡量投资回报率（ROI）。 [2:26] Yong Meng Sua: 它将系统设计为开放量子系统，把退相干和耗散作为动力学演化的‘燃料’而非阻碍。 [9:55] Yong Meng Sua: 去噪过度会丢失关键几何特征，去噪不足则信噪比极差 [9:55] Yong Meng Sua: 量子赋能而非取代经典 [14:04] Yong Meng Sua: 能否在本季度内实际运行并纳入业务预算？ [14:04] Yong Meng Sua: 能否与最顶尖的经典 CPU/GPU 集群进行硬碰硬的公开对标？ [14:04] Yong Meng Sua: 企业无需盲目投入数十亿美元，可以通过 90 天的试点项目（POC），从小切口算子开始严格对标 AI+量子的实际增益。

Source quotes

Yong Meng Sua: Quantum computing is often positioned as a future breakthrough, but meaningful value is already being realized today. Yong Meng Sua: These technologies are not replacements for classical AI, but powerful complements, particularly for complex optimization, pattern recognition, and data-intensive decision-making. Yong Meng Sua: Join this session to separate signal from noise and start leveraging quantum for practical impact today. [0:04] Yong Meng Sua: to the darker side of the world [0:04] Yong Meng Sua: When will a fault-tolerant quantum computer break RSA? [0:04] Yong Meng Sua: When will quantum computers finally compete with GPUs? [2:26] Yong Meng Sua: fuel [18:06] Yong Meng Sua: time and money are wasted

Original transcript

[0:04] Thank you, Peggy, for the introduction.  And thank you, the previous speaker, Nolan, [0:09] for such an inspiring story to begin our session  today. I cannot agree more that all the technology [0:16] created by human, whether it is AI or quantum, it  should serve for the greater good for the human. [0:24] Unfortunately, I have to bring you back to the  darker side of the world where I'm going to talk [0:27] about uh [sighs] from the hype to the real to  the utility making quantum useful for AI today. [0:40] So, as usual, I have to show you  this forward-looking statement [0:43] so that I don't I'm not getting into trouble  with the SEC later. So, let's get started with [0:51] the question they always get asked as someone  who working in quantum industry. So, typically, [0:56] you get asked the question that when will the a  fault-tolerant quantum computer will break the ISA [1:03] or when that when is that the moment that the your  quantum computer will be able to compete with GPU? [1:10] So, I think that's a wrong question to begin  with. So, there's a question that's always [1:15] anchor to the fact that you need to have maybe  millions of million of error corrected qubit. [1:20] You need to make sure that you have you have  to develop this in maybe 5, 15, 25 years [1:28] uh ahead in the future. And this is also where  that we always struggle to find a strategy to [1:35] deploy itself, right? So, I think the better  question to ask is what quantum can offer today? [1:43] For example, by using the quantum  photonics hardware that QC I is advocating. [1:48] So, I think it is fair for us that to  target a specific for example, to use [1:53] quantum photonics enable hardware to really  target a specific kernel for optimization for [2:00] temporal inferences or for a signal recovery. And  more importantly, all those things can be done [2:07] at room temperature. At today. And you can  measure this machine in a very very boring way, [2:14] in a way that you can compare with  a classical GPU CPU, for example. [2:18] So, you can compare that if it is worth your  time, your money to invest into this technology. [2:26] So, at QCI we believe in this vision. Today, we  have two offerings that we can bring to your hand, [2:33] whether to install in your premise or to use  it through the cloud subscription to machine. [2:40] The first one is the entropic quantum  optimization machine, which we call the tree [2:45] for the photonic analog solver. Where with this  machine we can actually have a quantum system [2:51] that use a dissipation to the for quantum  for vacuum fluctuation not in isolation [2:59] but actually to solve MP hard problem or  optimization problem toward the ground state, [3:04] which is your optimum solution. Another offering  we have today is photonic is a world computer, [3:10] where we actually using a neuromorphic base  inference so that we can encode the information [3:15] into the dynamic photonic state so they can use  the rich feature or non-linear feature in it [3:21] for the purpose of real time series prediction  or the signal level AI reconstruction. [3:27] And for that one, you can use it for pattern  recognition or edge compute or edge inference. [3:34] For this talk today, I'm going to take  a deep dive into our our offering of [3:38] entropic quantum computing. Where we use the  decoherence of the photonic state as a fuel [3:44] in a properly engineered way and not as barrier. [3:49] So this is the inversion to the gate-based quantum  computing. Where for the gate-based quantum [3:55] quantum computing, you always need to isolate  the the pristine photonic qubit from environment [4:02] so that you can use it without any error  correction. But for entropic quantum computing, [4:09] you use it as the open quantum system and you  can engineer the coupling of the system to the [4:14] environment so they can get best out of it. So,  what you can do is you can encode your problem. [4:22] You can actually put your problem into this  optical feedback loop and mix it with the [4:27] quantum quantum vacuum fluctuation. And then you  can let the system evolve to reach to the optimal [4:33] solution. And this is the first iterative. You can  keep the feedback into the system and then let it [4:39] run, for example, for 25, 50, 100 rounds so that  you you get to the statistical solution that you [4:44] need to assure that you reach a distribution of  the optimal solution. So, what we can have today [4:50] for Dark Tree. So, we're able to solve very  cluster problem. We're able to solve the NP-hard [4:57] problem, whether it is in discrete or continuous  optimization, in Kubo, in integer solver, [5:03] or analog solver through the qubit or multi-level  encoding. And you can have these in a very small [5:09] form factor, effective 5U form factor. We can  You can put it into a rack-mount photonics server [5:16] operating at room temperature and with  sufficiently low power. And you can access [5:20] this either the entire machine physically or to  access it through the cloud APIs subscription. [5:29] And you can fit in all AI stack, as I'm going to  jump into it later, where you can have a drop-in [5:34] solver for scheduling, for allocation, portfolio  optimization, or you can actually optimize sensor [5:40] placement or inference time decision layer. So,  the first model that I'm going to discuss today is [5:46] a diffusion sampling model where we call a quantum  reverse kernel. It is a programmable reverse [5:52] kernel that we can sample on the directory. So, a  diffusion model is a standard architecture where [5:57] you can generate where anyone can generate image  or video through it. So, what it does is started [6:06] from a neural network and then it will try to find  a solution iteratively. For example, another 25, [6:13] 50, 100 and then reach to a solution. As you can  see on the top layer of those reconstructed image. [6:20] For the QRK, we have a directory assembler  layer in this slide. We did We We can do it very [6:27] differently. In this case, we're not having just a  diffusion model, but we're having a directory as a [6:34] sampler in this diffusion model. So, in this case,  you actually can encode all the possible solution [6:40] in in this model and then ask the directory  to sample it parallelly at the same time. [6:46] So, what it does, it it will access all the  possible solution whenever you have this iterative [6:51] uh trial to find the optimum solution. So, we  can see that because of this unique structure, [6:57] you would do not have to retrain the  system. You use the same neural weight, [7:02] you same you use the same weight that you train,  and then you can just put insert a directory as [7:07] a sampling layer inside the iterative loop or  genetic loop. And you do not need an optimization [7:13] uh solver bolted on the side. And with these,  we can achieve a 46% better on a FID score, [7:21] which show that you can reach closer to  the ground truth. You can achieve about [7:27] 61% inception, which means that you can  attend a clearer sample as you can see here, [7:33] and then at the same time reconstruct  with better detail in the image. [7:39] In the second model where the quantum can  be useful into the air model is what we call [7:46] the discrete diffusion. So, this is in contrast  with what the model that we see before. Instead of [7:51] generating a pixel, what this model generates  is a series of data or sequences similar to the [7:58] DNA sequences. So, instead of generating a pixel  that's showing the for example, just now the MNIST [8:07] handwritten digit uh digit database, in this case  we break all the pixel into series or sequences. [8:15] Where we actually used instead of using a  mainstream sequence model that factorized [8:20] to get to the right solution, the D-Wave as an  internal solver in this model allow us to actually [8:28] use a native cubic solver into this model,  which means that we do not use a mathematical [8:34] model to try to reconstruct the correlation more  than just the pairwise data in the sequences. [8:42] Where we can actually use the high-order  correlation that naturally can be solved by D-Wave [8:47] to identify the correlation at high energy level  or high high-order correlation level to find a [8:52] better solution of the sequences or eventually the  picture. And with that, we're able to show that we [9:00] we always able to find a valid sequences that  can that reflect the ground truth with much [9:06] higher percentage, which is 89% versus 44% for  any standard mathematical mathematical model. [9:12] And if you draw a 15 candidate at a time, you  always find that one of those sequences that [9:19] you draw will reflect the ground truth. And  then with these, we're able always able to [9:24] generate the sequences that satisfy the triadic  motive rule once the cubic term is active in our [9:30] D-Wave inner solver. So, this shows that in a lot  of these image reconstruction or image generation, [9:36] the key is really to address the trial to come  in the payload. If you're able to identify higher [9:42] order correlation of the sequences of any data in  general, you're able to reconstruct a better image [9:48] or reconstruct better data that reflect the ground  truth. So, with all those models that I show you, [9:55] effectively speaking, only play with toy model.  How about if you use this with a real data that [10:00] we get from the real observation? So, in  the work that we have with NASA and the [10:06] for the last five project, what we learned is  they've been struggling in the past by using AI [10:13] to denoise the data, where they hit multiple  data bottleneck, whether it is for the denoising, [10:19] optimization, or corre- or correspondence. As all  the noise noisy data that they need to uh denoise, [10:26] they always come to the fact that if you denoise  too much, you lose the feature. You denoise too [10:32] little, you don't have sufficient signal. So, what  we learned here is if we add another quantum layer [10:38] or direct trace as solver into this AI model, you  can always construct a better data out. So, this [10:44] is one example where we denoise uh a with a very  weak lidar return amid strong sunlight background. [10:52] Direct trace in combination of AI model were  able to model the background noise and actually [10:57] use the correlation that we can mimic  on those noise to remove themself. So, [11:02] this is how we able to re- effectively remove the  noise without at the expense of the signal itself. [11:12] How about the deployment? So, one of the most  difficult challenge of any quantum modality, [11:18] whether you're using trapped ion, neutral atom,  photonic, or superconducting qubit. So, one of the [11:26] one of the biggest challenge is how you deploy  out from a lab. Right? You can have everything [11:32] nicely demonstrated in the lab, but if you don't  find a way to bring it to the hand of customer [11:38] with the right cost structure, there's no way you  can actually bring the device to be to to to get [11:44] the to to generate the greater good for the human  being. So, what we show here is we able to install [11:51] the dark tree in a commercial data center. So,  we have two example. One we install the dark tree [11:57] uh at the Quantum Corridor Quantum Safe Network in  uh Indiana. And the other one is we install at the [12:04] Forge at the at Central Federal Service Research  in Washington, D.C. So, this is one example where [12:08] you can see that we fit the dark tree into  the red color that rectangular box here into a [12:13] rack of the data center. And we did when we have  the dark tree sitting in this rack, we have a one [12:20] of our collaborator from the Tri for the Center  for Tri Institutional Therapeutic Therapeutic [12:27] Discovery Institute on using these exact model  to solve their protein folding problem. Where in [12:33] this case they can either define the more the most  optimal solution for the protein folding structure [12:39] for the purpose of drug creation. So, this show us  a very feasible way to really bring the technology [12:49] the photonic quant quant quantum photon hardware  out from the lab to the data center and eventually [12:54] fit into the any AI model potentially within  the right infrastructure. So, what does it mean? [13:03] Does it mean that we quantum will replace the  AI? I believe at this moment answer is always no. [13:09] We believe that IQC quantum is here to complement  to work hand in hand with uh GPU with any [13:17] AI the model. For example, we can always keep the  sensing injection part at the classical. You can [13:23] put this signal, denoise it with our quantum  plus AI solver and then you can always train [13:30] the air model that you need. Whatever it is, you  can train, you can infer it using the GPU and then [13:35] use another quantum layer to help it to decide, to  optimize and it will help you to act at the edge. [13:42] But the key is how to integrate all those  different stack layer in a very efficient way. [13:48] We believe that the best way to do that is  literally invest into the right architecture [13:53] that enable us to put quantum, for  example, any quantum processing unit [13:57] next to a GPU with the right within  the right structure or infrastructure. [14:04] But for us to distinguish the hype from the  utility, I think there are five questions [14:09] that we always need to answer. First of all,  this can I actually run this within this quarter [14:14] so that you can budget it, you can put it into  your business proposal? Can I actually have this [14:21] benchmark against the best classical machine?  I think this is very important. If you can use [14:25] a CPU or GPU to get the better result, I don't  see any reason you should use quantum at all. [14:32] And that is solve the problem you actually  have and the problem that matter can for [14:36] you to bring the value out. Eventually, it's  also important that we need to make sure that [14:41] how are we able to integrate and to have the  energy economic? Only with the right integration [14:48] and the right energy economic, we are able to  put the AI plus quantum model with the right [14:55] cost structure eventually that enable  the the revenue of the customer to scale. [15:02] So this is why Q-Ctrl is making  a strategic bet where we've been [15:07] through the recent acquisition that we  have on the set and half semiconductor. [15:11] We started the development of the 3D integration  of the thin film lithium niobate, a material [15:16] that we've been working for the last one and a  half decade to be fully integrated with the its [15:20] silicon photonics architecture. So, this will  allow us to put the thin film lithium niobate [15:26] wafer and multiple stack layer together with  silicon photonics and we can access all the [15:31] way down to the CMOS electronic. So, this  means that you're going to have the electro [15:36] uh the electrical and photonic interconnect in all  those three layer all the way from the CMOS layer [15:42] to the thin film lithium niobate in a multiple  complex structure. So, why we cannot use either [15:49] only thin film lithium niobate or silicon  photonics platform alone? Because we need both [15:53] to solve the larger scale problem. We need an  unmatched electro-optical and also nonlinear [15:58] optics physics of the thin film lithium niobate.  At the same time, we don't want to actually [16:03] dispose the all the architecture that have been  built on this silicon photonic infrastructure [16:10] enabled by the semiconductor industry. So, only by  the combination of both we can achieve the energy [16:16] energy per bit for us to have all the AI binding  constraint, we can actually have the product [16:23] can get benefited from the existing  semiconductor semiconductor ecosystem. [16:28] And at the same time, we can have a fully  integrated supply chain within United States. [16:35] And to do that, we do not need to  invest billions of dollars to that. [16:40] You can simply take a 90-days pilot project  with some cost fund with some quantum members [16:45] such as QCRI. I think you start small, find  a kernel that work for your AI model and then [16:51] benchmark honestly what is the improvement that  you can get from this AI plus quantum model. [16:57] And then from there, if you see positive result,  the next step is how to integrate hybridly [17:01] so that it can work for you to create value for  you or your customer in the right cost structure. [17:09] So, if there's only three things that you can take  from this talk, I think that today, I hope this [17:15] in res- what in resonance with you is the utility  doesn't wait for the photon and quantum computer. [17:21] We don't need that. In fact, I think it is always  best to find a way for quantum to work with AI so [17:28] they can get the values become the early adopter  in this race. I would believe that quantum [17:33] will complement AI and doesn't compete with that  because we see that this is how we can bet get the [17:38] best out quantum at the same time and able quantum  to address the shortcoming of the AI itself. [17:47] And I think the lastly we have to be very honest.  We have to judge on the performance of the [17:52] any AI model whether it is a pure AI model, AI  plus quantum model to to to to look in clearly [17:59] whether this is something that bring value to  you, not just the promises 20 20 years from now. [18:06] Thank you. That's all for me. Um I got one  question for you. So um from the model for [18:12] your hardware looking how these things going to  be and it looks like a fairly complex uh system. [18:17] How cost effective it is right now for industry  build something making to its application? [18:24] There's a very fair question. In fact,  what I showed you just now is if I add [18:29] a directory solver into my AI model, I always able  to see some improvement based on the if I compare [18:35] and benchmark again a quantum solver. But what  I didn't show you are the shortcoming of this is [18:41] whenever I need to have the data stream to my  directory the the physical time of have having [18:48] this transition is long. I need to get from the  client side software to the embedded software and [18:53] then get into the system and then vice versa.  So that is the there is really where the time [18:58] and the money wasted. At today, if you want to  scale, I think we need to find a better way to [19:03] scale. This is why Q-CTRL is invested into the  3D integration photonic chip so that you can have [19:10] a a chip with a similar architecture similar to  the photon in the connector being used in the [19:14] data center so they can have your photonic chip  on the photonic hardware core package with GPU [19:20] so that you don't have all those  letters that bring the cost up.