Why new AI hardware has to prove itself in a real system
Lumai’s optical Iris servers are being deployed at ARIA’s Scaling Inference Lab. Dr Rosemary Francis explains in an interview why this matters.
Lumai was selected through a competitive technical evaluation to join the Scaling Inference Lab. The lab is run by CommonAI CIC with £50 million from ARIA, the UK’s Advanced Research and Invention Agency. CommonAI Compute is a member of the lab.
Lumai’s Iris servers are being deployed across multiple nodes in a commercial data centre. Iris uses light to do the matrix maths behind prefill, the compute-heavy phase where an AI model processes the prompt. GPUs handle decode, where the model generates its answer. The lab will test deployability, reliability and disaggregated inference: optical prefill and silicon decode working together in one stack.
Why it matters: three points from the interview
- AI growth is increasingly limited by power.
- No single chip is right for every part of an inference workload.
- New hardware must prove itself in a complete system, on real workloads.
Rosemary spoke about choice and power:
“We need to make sure that we get different choices where we get our compute from, and power is a huge part of it. We’ve not invested in the UK in our power infrastructure in a way other nations have. And now we have additional power constraints for the AI technologies that we want to build.”
Lumai’s CEO, Dr Xianxin Guo, put it simply: “Optical computing is not just a concept, it’s real.”
Watch the film: ARIA, CommonAI and Callosum on why this deployment matters.