中心博士生易思成的工作——Global-Power-Split-Tree Architecture for Large-Scale Coherent Optical Matrix Multiplication（面向大规模相干光计算的二叉树全局分光架构）的相关成果近期被PHYSICAL REVIEW APPLIED期刊接收发表，该工作得到了国家自然科学基金（T2225023，62205203）的部分资助。光子学在物理上保证了矩阵乘法的高速和低消耗计算。然而，光子矩阵乘法的可扩展性受到系统级联光学移相器插入损耗及耦合器分光损耗的强烈限制，很难在单一架构中同时实现大吞吐量、高精度、低功耗和高密度。在这里，我们提出了一种二叉树全局分光架构，它以二叉树的形式部署光学移相器，降低器件插损及分光总损耗，显著提高了可扩展性。通过理论分析及仿真结果表明，二叉树全局分光架构更适应于光学矩阵架构大规模化，相对于传统架构，集成规模可提升约100倍，能效比可提升约10倍。进一步采用5种常用的神经网络对本架构进行通用性测试，我们发现与传统架构相比，本架构能耗降低近10倍。
摘要: Photonics holds the physical potential for achieving both high-speed and low-consumption matrix multiplication. Nonetheless, due to the insertion loss of optical phase shifters and the loss incurred from power splitting, traditional methods are challenging to achieve high throughput, precision, and energy efficiency within a single framework. In this letter, we propose a global-power-split-tree (GPST) architecture for large-scale coherent optical matrix multiplication. The insertion loss of phase shifters decreases from O(N ) to O(logN ) when utilizing phase shifters in the form of splittree configurations. The architecture enhances energy utilization from O(1/N2) to O(1/N ) through global power allocation. Theoretical analysis and simulation show that, compared with conventional coherent optical matrix multiplication architectures, GPST can achieve near 10× higher energy efficiency. By employing five different neural networks to illustrate the overall power consumption of GPST, we find that compared to conventional architectures, power consumption decreases by an order of magnitude, especially when the integration scale is large, compared to conventional architectures.