Photonic Tensor Processing Chip PhotonCore Released: Optical Computing Surpasses Traditional GPUs in AI Inference Tasks for the First Time

Photonic Tensor Processing Chip PhotonCore Released: Optical Computing Surpasses Traditional GPUs in AI Inference Tasks for the First Time

Optical computing company Lightmatter released the commercial photonic tensor processing chip PhotonCore on August 24, 2030. In standardized AI inference task tests, PhotonCore achieved 47 times the inference performance per watt of the NVIDIA H200 GPU and 12 times the throughput, marking the first time optical computing technology has fully surpassed traditional electronic computing in practical applications.

PhotonCore uses silicon photonics integration technology, integrating over 10,000 optical matrix multiplication units on a single chip. Unlike traditional GPUs that require repeated data movement between memory and compute units, PhotonCore leverages the superposition and interference properties of light to perform matrix operations directly in the optical domain, fundamentally eliminating the data movement bottleneck.

Lightmatter CEO Nicholas Harris demonstrated PhotonCore running the GPT-7 model at the launch event. On the same inference task, PhotonCore's response latency was 12 milliseconds compared to 156 milliseconds for the H200. Even more significant was the power consumption difference: PhotonCore's full-load power consumption is just 45 watts, compared to 700 watts for the H200.

"Optical computing is not a replacement for electronic computing — it is an entirely new computing paradigm for the AI era," Harris said. "When AI model parameters exceed the trillion level, electronic computing's power consumption and heat dissipation will become insurmountable physical limits, while optical computing is inherently suited for massively parallel computation."

PhotonCore's technical breakthrough lies in "programmable optical phase-change materials." Traditional optical computing chips have optical elements that cannot be altered once manufactured, whereas PhotonCore uses phase-change material (GST) as optical modulators, capable of changing the refractive index of optical elements at nanosecond speeds, enabling programmable optical matrix operations.

In actual data center deployment tests, Google Cloud used PhotonCore clusters to run large-scale AI inference services, reducing power consumption by 89% at equivalent compute capacity. Google Cloud's VP of Infrastructure said: "PhotonCore not only reduces costs — more importantly, it makes AI services that were previously constrained by power and heat dissipation feasible."

However, PhotonCore currently only supports inference tasks and does not support AI model training. Lightmatter's Chief Scientist explained: "Training requires high-precision gradient computation, and the analog nature of optical computing still has precision limitations. We expect optical training chips to launch by 2032."

NVIDIA responded that its next-generation GPU architecture Blackwell Ultra will launch in 2031, narrowing the energy efficiency gap with optical computing through chiplet packaging and advanced cooling technology. However, analysts believe optical computing's advantage in the inference domain is already irreversible, and traditional GPU manufacturers need to accelerate their transition to optical computing.

PhotonCore's developer edition opened for pre-order on August 25, priced at $12,000 per chip. Lightmatter expects 2031 production capacity to reach 100,000 units, primarily supplied to hyperscale data center operators.