UK firm unveils Iris Nova, the world's first optical system for real-time LLM inference, reducing power by 90%.
Moving LLM inference from silicon to optical computing addresses the fundamental power wall bottlenecking AI scaling. If Iris Nova's 90% power reduction holds in production workloads, it completely changes the unit economics of hosting large models. The key test will be integration overhead and whether the analog-to-digital conversion negates the latency benefits.
The UK-based hardware startup behind Iris Nova has introduced what it claims to be the world's first optical computing system designed specifically for real-time Large Language Model (LLM) inference. By utilizing light rather than electrical currents to perform the massive matrix multiplications required by neural networks, the Iris Nova server reportedly consumes up to 90% less power than traditional silicon-based GPU clusters.
Technical Context Standard silicon accelerators are increasingly constrained by the power wall and thermal dissipation limits. As transistor scaling slows, pushing more current through silicon generates unsustainable heat. Optical computing bypasses this by using silicon photonics. In systems like Iris Nova, data is encoded into light pulses, and calculations (such as multiply-accumulate operations) occur passively as light passes through optical interference networks. This allows computations to happen at the speed of light with near-zero energy loss to heat.
Why It Matters From an engineering perspective, inference is a fundamentally different workload than training; it is highly sensitive to latency and operational expenditure (OpEx). Power consumption is currently the dominant OpEx factor for AI services. If a 90% reduction in power is achievable at scale, it fundamentally alters the unit economics of deploying LLMs. Furthermore, freeing up power budgets allows data centers to deploy higher densities of compute in power-constrained regions, breaking through the physical limits of current data center infrastructure.
What to Watch Next While the theoretical benefits of photonic computing are well-documented, the practical challenges lie in the analog-to-digital conversion (ADC) and digital-to-analog conversion (DAC) steps at the edges of the optical chip. Watch for independent benchmarks detailing the end-to-end latency, specifically whether the I/O overhead negates the optical processing speed. Additionally, engineers should monitor the system's precision capabilities—optical computing is inherently analog and can introduce noise, so seeing how Iris Nova maintains model accuracy (e.g., FP16 or INT8 equivalence) on standard open-weight models will be the true test of its viability.