ARIA 选择用于 Rack 比例 AI 推论测试的 Lumai 光学服务器

AI 模型和平台

ARIA 选择用于 Rack 比例 AI 推论测试的 Lumai 光学服务器

在 Google 上的首选来源中添加 Unite.AI

Lumai将光学AI硬件带入一个新的评价环境：由英国高级研究和发明机构(ARIA)支持的扩展推论实验室。 实验室计划部署Lumai Iris推论服务器，作为其第一个光学计算平台，为研究人员提供了一个在商业数据中心环境中在架尺度上测试架构的方法。

由CommonAI CIC运行，实验室旨在为新兴AI硬件弥补一个困难的缺口：演示一个新颖的处理器是一回事；使它作为一个集成系统的一部分发挥作用，具有真实的模型和普通的动力，冷却，和机架要求是另一回事。

公告以伊里斯·诺瓦的4月发布为基础。 新步骤是通过一个公开支持的测试台进行多节点评价，而不是发布一个完整，独立的基准测试来证明卢迈的头条效率声称。

使用光执行矩阵数学

卢迈(英語：Lumai)是一所成立于2021年的牛津大学旋转式建筑，开发了混合光学和数字建筑。 其光学引擎针对AI推论中使用的密集数学操作，而电子组件则继续处理周围的系统功能。

在公司的技术描述中，信息被编码成光束，并通过三维空间的光学组件处理。 其目的是通过光传播平行地执行许多数学操作，而不是通过切换晶体管来进行每一个计算。

这并不能使整个服务器被动或无电。 完整的系统仍然需要光源，数据转换，内存，数字处理，以及控制。 相关的工程问题是，当一个有用的模型实际运行时，光学发动机的效率在多大程度上能维持这些周围的成本。

Lumai描述了一个广的数据接口，旨在将转换管理费分散到庞大的工作量中，一个用于模型的键值缓存的三级内存设计，以及硬件意识的量化。 这些是结构的实用部分：移动数据、保留可用数字精度、以及使加速器提供工作都影响实验室演示以外的性能。

为什么卢迈是目标预填

大语言模型推论有两个不同的阶段。 在预填过程中，系统处理输入并构建生成响应所需的内部状态。 在解码过程中，它一步一步地产生输出符，反复访问模型权重和缓存上下文。

预充往往以大型矩阵操作为主，而解码则会受到内存带宽的限制，特别是在较小的批量尺寸时。 平衡随模型大小，即时长度，批量，以及服务配置而变化。 卢迈的方法是专门从事计算重的阶段，而不是用一种处理器优化整个管道。

它的Iris产品文档将光学加速器定位在多样的计算系统内，包括作为预填处理器与优化解码的硬件一起使用。 Lumai 称 Iris Nova支持Llama模型，有80亿和700亿参数，可用于数据中心评价。 产品家族设计为标准架式基础设施和空气冷却。

称 Xianxin Guo, Lumai的共同创始人兼首席执行官。 ARIA的工作将整合问题置于下一阶段的中心。

完整系统测试台

ARIA是一家英国公共研发供资机构，它支持雄心勃勃的技术，具有潜在的社会效益。 它的扩展计算方案探索既定计算方法的替代品，包括新的计算原始、联网和系统设计。

扩展推论实验室有5 000万英镑的ARIA支持。 ARIA将其描述为一系列结构化的架子级飞行员，每个飞行员将未经证实的技术纳入一个可部署的系统，最终用户可以运行模型。 目标是将全系统的性能与商业替代品进行比较，并迅速进行迁移，每六个月计划一个新的架子。

CommunityAI，剑桥非营利会员组织提供实验室，带来共同公司，公共机构，合作工程的学术参与者。 其方案大纲描述了超过五个架级系统，并强调独立验证以帮助新兴硬件展示真正的操作性能。

这种结构很重要，因为一旦数据移动、调度、软件支持和其余服务管道都包括在内，加速器在孤立的操作上能够发挥良好的作用，同时提供较小的好处。 共享的试验环境有助于揭示优点和瓶颈。

效率声称做- 不做- 显示

Lumai 称架构有可能将数据中心功率消耗降低90%，并使拥有权的总成本达到常规GPU架构的十分之一左右。 其网站描述了50倍业绩的路线图。 当光学预填装与硅基解码配对时，单独报告每推算的能源效率至少提高16倍。

这些数字描述的是不同的比较，不应作为可互换的措施或已经由ARIA独立核实的结果处理。 该公告没有提供常见的基准测试配置，模型质量比较，或公布的架子级结果集，可以确定跨商业工作量的收益。

因此，实验室的评估是实质性的发展。 有用的测量将包括每个完成的请求的能量、第一信号的时间、持续的产出吞吐量、模型精度以及操作整个系统的费用。 进行有意义的比较需要同样的模型、上下文长度、批量假设和服务要求。

对于英国在有限的电力预算范围内扩大国内AI能力的雄心，成功的光学系统可以提供另一种计算选择。 卢迈的部署是在商业条件下检验这种可能性的一个步骤。 其优势是否转化为规模较低的服务费用将取决于评价所产生的全系统证据。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

AI Models & Platforms

ARIA Selects Lumai Optical Servers for Rack-Scale AI Inference Testing

Add Unite.AI to your preferred sources on Google

Lumai is taking its optical AI hardware into a new evaluation setting: the Scaling Inference Lab backed by the UK’s Advanced Research and Invention Agency, or ARIA. The lab plans to deploy Lumai Iris inference servers as its first optical-compute platform, giving researchers a way to test the architecture at rack scale in a commercial data-center environment.

Run by CommonAI CIC, the lab is intended to bridge a difficult gap for emerging AI hardware: demonstrating a novel processor is one thing; making it work as part of an integrated system, with real models and ordinary power, cooling, and rack requirements, is another.

The announcement builds on the April launch of Iris Nova. The new step is a multi-node evaluation through a publicly backed testbed, rather than the release of a completed, independent benchmark proving Lumai’s headline efficiency claims.

Using Light to Perform the Matrix Math

Lumai, an Oxford University spin-out founded in 2021, develops a hybrid optical and digital architecture. Its optical engine targets the dense mathematical operations used in AI inference, while electronic components continue handling the surrounding system functions.

In the company’s technical description, information is encoded into beams of light and processed through optical components in three-dimensional space. The aim is to execute many mathematical operations in parallel through light propagation, rather than perform every calculation by switching transistors.

That does not make the entire server passive or power-free. The full system still needs light sources, data conversion, memory, digital processing, and control. The relevant engineering question is how much of the optical engine’s efficiency survives those surrounding costs when a useful model is actually running.

Lumai describes a wide data interface intended to spread conversion overhead across large workloads, a three-tier memory design for the model’s key-value cache, and hardware-aware quantization. These are practical parts of the architecture: moving data, preserving usable numerical precision, and keeping the accelerator supplied with work all affect performance outside a laboratory demonstration.

Why Lumai Is Targeting Prefill

Large-language-model inference has two distinct phases. During prefill, the system processes the input and builds the internal state needed to generate a response. During decode, it produces output tokens step by step, repeatedly accessing model weights and cached context.

Prefill is often dominated by large matrix operations, while decoding can be constrained by memory bandwidth, particularly at smaller batch sizes. The balance changes with model size, prompt length, batching, and serving configuration. Lumai’s approach is to specialize in the compute-heavy stage rather than optimize the entire pipeline with one type of processor.

Its Iris product documentation positions the optical accelerator within heterogeneous computing systems, including use as a prefill processor alongside hardware optimized for decode. Lumai says Iris Nova supports Llama models with 8 billion and 70 billion parameters and is available for data-center evaluation. The product family is designed for standard rack infrastructure and air cooling.

“Running a multi-node system in a commercial data center is where new compute architectures prove whether they are ready for real deployment,” said Xianxin Guo, Lumai’s co-founder and CEO. The ARIA work puts that integration question at the center of the next phase.

A Testbed for Complete Systems

ARIA is a UK public research-and-development funding agency that backs ambitious technologies with potential societal benefits. Its Scaling Compute programme explores alternatives to established computing approaches, including new computational primitives, networking, and system design.

The Scaling Inference Lab has £50 million in ARIA backing. ARIA describes it as a series of structured rack-level pilots, each incorporating an unproven technology into a deployable system that end users can run models on. The goal is to compare full-system performance with commercial alternatives and iterate rapidly, with a new rack planned every six months.

CommonAI, the Cambridge-based nonprofit membership organization delivering the lab, brings together companies, public bodies, and academic participants for collaborative engineering. Its programme outline describes more than five rack-level systems and emphasizes independent validation to help emerging hardware demonstrate real operating performance.

That structure matters because an accelerator can perform well on an isolated operation while delivering a smaller benefit once data movement, scheduling, software support, and the rest of the serving pipeline are included. A shared test environment can help expose both the advantages and the bottlenecks.

What the Efficiency Claims Do—and Do Not—Show

Lumai says the architecture has the potential to reduce data-center power consumption by up to 90% and bring total cost of ownership to around one-tenth of conventional GPU architectures. Its website describes a roadmap toward 50 times higher performance. The release separately reports at least 16 times better energy efficiency per inference when optical prefill is paired with silicon-based decode.

These figures describe different comparisons and should not be treated as interchangeable measures or as results already independently verified by ARIA. The announcement does not provide a common benchmark configuration, model-quality comparison, or a published rack-level results set that would establish the gains across commercial workloads.

The lab’s evaluation is therefore the substantive development. Useful measurements would include energy per completed request, time to first token, sustained output throughput, model accuracy, and the cost of operating the complete system. The same model, context lengths, batching assumptions, and service requirements are needed for meaningful comparisons.

For the UK’s ambition to expand domestic AI capacity within limited power budgets, a successful optical system could offer another computing option. Lumai’s deployment is a step toward testing that possibility under commercial conditions. Whether its advantage translates into lower serving costs at scale will depend on the full-system evidence the evaluation produces.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.