This Optical Computing Chip Achieves 10x Nvidia’s Performance—Can It Break Silicon’s Limitations?

AI chip company Neurophos, based in Austin, Texas, recently announced breakthrough test results for its optical processing unit (OPU). The company claims that its self-developed OPU delivers ten times the computing power of Nvidia’s latest Vera Rubin NVL72 AI supercomputer under FP4/INT4 workloads, while maintaining similar power consumption.

According to CEO Patrick Bowen, the performance leap comes from two core innovations: ultra-large computing matrices and high operational clock frequency. The OPU integrates a 1000×1000 pixel photonic sensor matrix, over 15 times larger than the 256×256 matrices used by most AI GPUs. Meanwhile, the physical size of its phototransistors has been reduced by roughly 10,000×, solving the integration bottleneck of traditional silicon photonic chips.

Hardware-wise, the Tulkas T100’s core unit occupies roughly 25 mm², equivalent to a single tensor core in “optical terms.” Although it has fewer cores than Nvidia’s 576 tensor-core Vera Rubin chip, the combination of large matrices and 56 GHz clock frequency enables it to surpass Nvidia GPUs in actual computing performance. The clock speed is six times higher than the record 9.1 GHz of Intel Core i9-14900KF and far exceeds Nvidia RTX Pro 6000’s 2.6 GHz.

Importantly, Neurophos’s phototransistor technology is compatible with existing semiconductor manufacturing systems, paving the way for potential collaborations with major fabs like Intel and TSMC. The chip is currently in laboratory testing, with mass production expected by 2028. Key challenges remain, including vector processing unit integration and synchronization between SRAM and optical computing units.

Photonics, with its high speed and low power consumption, has become a major focus for global tech giants. Nvidia’s Vera Rubin already integrates the Spectrum-X optical Ethernet switch, while AMD plans to invest $280 million to build a silicon photonics R&D center. Numerous startups are also entering the field, diversifying optical computing research.

Neurophos’s breakthrough demonstrates the natural advantages of optical computing in AI model training and low-precision inference. By replacing electrons with photons, it physically overcomes the heating and frequency limits of traditional chips. Although challenges remain in mass production yield, cost, and electronic architecture compatibility, photonics offers a promising path to breaking silicon’s computing limits.

As global AI demand continues to surge, photonics may become the core of next-generation AI computing. Neurophos’s progress, alongside Nvidia and AMD’s investments, will accelerate photonics adoption. Heterogeneous photon-electron architectures could become the mainstream for future AI supercomputers.