AI计算或迎来“光的时代”

Core Viewpoint - The company Guangbenwei Technology is developing glass-based optical computing chips to replace silicon substrates, aiming to achieve a new era of "1000 tops computing power and 1000 tops/W energy efficiency" for AI applications [1][2]. Group 1: Technology and Innovation - The glass optical computing chips are designed for AI inference scenarios, with the market expected to reach $255 billion by 2030, as inference will account for 75% of total AI computing [1]. - The company has completed its first commercial-grade optical computing chip, with a matrix size of 128mm x 128mm and a peak computing power exceeding 1000 tops [1]. - Glass is chosen over silicon due to its superior properties such as flatness, thermal stability, and compatibility with optical waveguide processes, allowing for larger chip designs and improved performance [2]. Group 2: Performance and Efficiency - The glass optical computing chip can achieve a computing power of 2600 tops when sized at 200mm x 200mm, significantly enhancing computational capabilities [2]. - The company has developed a zero static power consumption feature for its optical computing chips, predicting an energy efficiency ratio exceeding 1000 tops/W, which is over 200 times that of TPU [2]. Group 3: Future Developments - Guangbenwei Technology aims to create a next-generation all-optical computing system that allows light signals to perform repeated calculations and dynamic storage, moving beyond the limitations of single-core optical computing [3]. - The ultimate goal is to package the glass optical computing chip into a high-performance all-optical computing system, defining the next generation of AI computing standards and providing comprehensive solutions for various user scenarios [3].