Large-scale photonic chiplet Taichi empowers 160-TOPS/W artificial general intelligence

• DOI: 10.1126/science.adl1203
• Publication Date: 2024-04-11T17:58:47Z
• Authors: Zhihao Xu; Tiankuang Zhou; Muzhou Ma; ChenChen Deng; Qionghai Dai; Lu Fang
• Keywords: TOPS; Tera-

The pursuit of artificial general intelligence (AGI) continuously demands higher computing performance. Despite the superior processing speed and efficiency integrated photonic circuits, their capacity scalability are restricted by unavoidable errors, such that only simple tasks shallow models realized. To support modern AGIs, we designed Taichi-large-scale chiplets based on an diffractive-interference hybrid design a distributed architecture has millions-of-neurons capability with 160-tera-operations per second watt (TOPS/W) energy efficiency. Taichi experimentally achieved on-chip 1000-category-level classification (testing at 91.89% accuracy in 1623-category Omniglot dataset) high-fidelity intelligence-generated content up to two orders magnitude improvement paves way for large-scale advanced tasks, further exploiting flexibility potential photonics AGI.