Optical Computing

This pilot aims to deploy the world’s first rack-scale optical AI inference cluster, pairing eight Lumai Iris Nova optical inference servers with NVIDIA GPUs in a disaggregated architecture. Optical hardware takes on the compute-bound prefill stage while GPUs handle memory-bound decode. The cluster explores whether running matrix multiplications in light rather than in electronic circuits can deliver substantial gains in performance per watt for end-to-end LLM inference.