英国高级研究与发明局（ARIA）资助的扩展推理实验室将部署Lumai的Iris服务器，作为其首个光计算平台。Lumai宣称可大幅节省能源，但这些数据尚未经过独立测试。该实验室将在商业数据中心对这些服务器进行测试。
ARIA’s Scaling Inference Lab will deploy Lumai’s Iris servers, which use beams of light to run AI calculations, as its first optical platform. Lumai claims big energy savings, but these have not been independently tested. The lab will test the servers in a commercial data centre.
牛津初创公司Lumai周二表示，英国高级研究与发明局（ARIA）将在其资助的一个实验室里，测试利用光运行AI模型的服务器。
The UK’s Advanced Research and Invention Agency (ARIA) will test servers that use light to run AI models, in a lab it funds, Oxford startup Lumai said on Tuesday.
由非营利组织CommonAI运营、获得ARIA 5000万英镑资助的扩展推理实验室，将安装Lumai的Iris服务器作为其首个光计算平台。实验室将在商业数据中心跨多台机器运行这些服务器，以观察它们在真实AI模型、电力和散热条件下的表现。
The Scaling Inference Lab, backed by £50M from ARIA and run by the nonprofit CommonAI, will install Lumai’s Iris servers as its first optical computing platform. It will run them across several machines in a commercial data centre, to see how they cope with real AI models, power and cooling.
Lumai于2021年从牛津大学分拆而出，将数据编码进光束。其芯片利用光而非晶体管开关完成AI背后的繁重数学运算，而普通电子元件处理其余部分。同样基于牛津研究的Salience Labs上月为光交换机融资1.25亿美元。
Lumai, spun out of Oxford University in 2021, encodes data into beams of light. Its chips do the heavy maths behind AI using light rather than switching transistors, while normal electronics handle the rest. Salience Labs, which also builds on Oxford research, raised $125M last month for optical switches.
该公司将硬件瞄准“预填充”阶段，即模型首次读取用户提示词的阶段。普通芯片仍将负责生成答案。
The company aims its hardware at “prefill”, the stage where a model first reads a user’s prompt. Ordinary chips would still produce the answer.
“Lumai的技术是我们评估过的最成熟的光计算平台，”ARIA首席科学与技术官David Lewis表示。
“Lumai’s technology is the most mature optical compute platform we evaluated,” said David Lewis, ARIA’s chief science and technology officer.
Lumai宣称其方法可将数据中心能耗降低多达90%。公司还称，当与标准芯片搭配时，其系统每请求能效至少是单纯硅基方案的16倍。这些数据均未经独立核实，验证它们正是实验室的任务。
Lumai claims its approach could cut data centre power use by up to 90%. It also says the system is at least 16 times more energy efficient per request than silicon alone, when paired with standard chips. None of these figures has been independently checked, and testing them is the lab’s job.
该实验室已引入英国初创公司Oriole，后者利用光互联AI芯片。与此同时，英伟达已在数据中心内部的光互连上投入数十亿美元。另一家英国推理芯片初创公司Fractile正按65亿美元估值融资。
The lab already hosts UK startup Oriole, which uses light to link AI chips. Nvidia, meanwhile, has spent billions on light-based connections inside data centres. Another UK inference-chip startup, Fractile, is raising at $6.5B.