Photonic Computing Chip PhotoCore Released: Replacing Electrons with Photons for 95% Energy Reduction

Dutch photonic computing startup LumiSilicon today launched PhotoCore in Amsterdam, the worlds first commercially available photonic computing chip. Unlike traditional electronic chips, PhotoCore uses photons to perform matrix operations, the most fundamental computation in AI inference.

Commercializing Photonic Computing

Photonic computing has existed as a concept for decades but remained confined to laboratories. LumiSilicon founder and CEO Maarten de Jonge stated that PhotoCores breakthrough lies in solving a key engineering challenge: integrating optical components into standard semiconductor manufacturing processes.

PhotoCore uses LumiSilicons proprietary silicon-photonic hybrid architecture. Electronic transistors on the chip handle control and scheduling logic, while photonic waveguide arrays are dedicated to massively parallel matrix operations. Photons propagating through waveguides complete multiply-accumulate operations through interference effects, generating almost no heat.

Performance Data

In benchmarks published by LumiSilicon, PhotoCore achieved an energy efficiency of 420 TOPS per watt when running large language model inference, approximately 20 times that of Nvidias B300. Latency for a standard Transformer inference task was 0.8 milliseconds versus 16 milliseconds for the B300.

However, de Jonge acknowledged that photonic computing currently excels at dense matrix operations. For sparse computation and branch-logic-intensive tasks, traditional electronic chips remain superior. PhotoCores training capability is also limited, as gradient backpropagation during training is difficult to implement in optical architectures.

First Customers and Pricing

LumiSilicon has partnered with Microsoft Azure and Alibaba Cloud, both of which will offer PhotoCore-based AI inference instances starting Q3 2030. Individual PhotoCore chips are priced at $12,000, competitive with the B300s $15,000. De Jonge expects costs to fall below $5,000 within two years as manufacturing matures. The launch marks a pivotal turning point for photonic computing transitioning from lab to commercial market.