Lumai says its optical AI servers will be onboarded for evaluation at the UK’s ARIA-backed Scaling Inference Lab.
The big change
- The proposed test would place Iris servers in a commercial data-centre environment.
- Lumai targets prefill, the stage that processes a model’s input before it answers.
- Rack-scale testing is planned; the company’s claimed power savings are not independently established.
ARIA describes the CommonAI-delivered lab as a testbed for rack-level systems. Partners contribute technology or resources, while ARIA funds integration and testing. The programme aims to compare complete systems with commercial alternatives.
Lumai says Iris encodes numbers in light intensity and uses passive optical components for mathematical operations. Its proposed architecture targets matrix-heavy prefill while pairing optical processing with silicon for decoding. The company describes a hybrid system rather than a data centre operating entirely on light.
NVIDIA’s general explanation separates inference into two phases: prefill processes input tokens and builds intermediate states used for the first output token; decoding generates subsequent tokens. The phases use resources differently, so an improvement in one stage alone does not establish an end-to-end gain across workloads.
In its November 3 announcement, Lumai says the approach could cut data-centre power consumption by up to 90%. It supplies no independent measurements, comparison conditions or specific workload for that claim.
