比尔盖茨押注硅光突破：旗下Neurophos首款光子芯片性能达英伟达AI超算十倍

Core Insights - Neurophos, an AI chip startup backed by the Gates Frontier Fund, has made significant advancements in silicon photonics, developing an optical processing unit (OPU) that is approximately 10,000 times smaller than existing technologies and capable of a 1000×1000 pixel scale photonic computing matrix on a single chip [1] Group 1: Technology Advancements - The first optical accelerator, Tulkas T100, achieves AI computing performance that is ten times greater than NVIDIA's latest Vera Rubin NVL72 supercomputer at FP4/INT4 precision, while maintaining similar power consumption levels [3] - Key technologies contributing to this performance include a 1000×1000 photonic tile, which significantly exceeds the mainstream GPU matrix size of 256×256, and a clock frequency of 56 GHz, which is much higher than the 9.1 GHz of Intel's Core i9-14900KF and 2.6 GHz of NVIDIA's RTX Pro 6000 GPU [3] Group 2: Manufacturing and Future Prospects - The CEO of Neurophos, Patrick Bowen, stated that traditional optical transistors produced in silicon photonic factories are about 2 millimeters long, making high-density integration challenging. Their technology miniaturizes optical transistors to a scale compatible with CMOS processes, enabling large-scale parallel optical computing [3] - The Tulkas T100 chip contains only one "optical tensor core" with an area of approximately 25 square millimeters, significantly fewer than the 576 digital tensor cores integrated into NVIDIA's Vera Rubin chip, yet achieves higher effective throughput through greater matrix dimensions and clock frequency [3] - Despite these advancements, Neurophos acknowledges that the technology is still in the engineering validation stage, with mass production not expected before 2028, and several challenges remain, including on-chip SRAM capacity, vector processing unit expansion, and optoelectronic co-design [4]