Lumai Iris servers will join a UK research laboratory to test optical computing on real artificial intelligence workloads. CommonAI will run the evaluation in a commercial data centre. The design assigns prompt processing to light-based hardware and response generation to conventional chips; the programme will evaluate the combined system under ordinary power, cooling and rack conditions.
The United Kingdom’s Advanced Research and Invention Agency, ARIA, will put Lumai Iris servers into its Scaling Inference Lab for an evaluation of optical computing. The laboratory is operated by CommonAI, a nonprofit working on collaborative artificial intelligence. Its planned tests will use real model workloads across multiple machines in a commercial data centre, with ordinary rack, power and cooling arrangements.[1], [2]
Lumai, a computing company spun out of the University of Oxford, represents numbers through the intensity of light beams. Optical components perform the matrix calculations used by language models. Iris targets prefill: the calculation-heavy stage in which a model processes the prompt before producing its first answer token. Conventional silicon chips remain responsible for decode, the stage that generates the response. The proposed system combines those two types of hardware.[1], [2]
Iris will be the laboratory’s first optical inference platform. ARIA’s chief science and technology officer David Lewis described the lab as a setting for open, independent testing at rack scale. The Next Web reports that Lumai’s efficiency estimates have not received independent verification. The announced collaboration is an evaluation programme, with the commercial installation intended to test the technology’s operation under data-centre conditions.[2]
