65 TOPS optoelectronic multi-core computing unlocking multi-feature fusion enhancement

Abstract

The rapid progress of artificial intelligence demands scalable, energy-efficient hardware with massive parallelism and high throughput. Although optical computing offers a promising post-Moore solution, current implementations face speed limitations and multi-core integration challenges. Here, we propose a high-throughput optical processing unit (OPU) that simultaneously exploits coherent interference, wavelength-division multiplexing, and spatial parallelism to achieve revolutionary performance gains. Integrating four optical analog cores on a monolithic chip, the OPU supports 124-channel parallel task processing, achieving 65.04 trillion operations per second (TOPS) computational speed and 5.16 TOPS/mm2 compute density. Leveraging this OPU platform, an optoelectronic convolutional neural network (OE-CNN) is constructed that fuses the OPU’s parallel 4-kernel convolution and average pooling operations with electronic nonlinear activation and fully connection operations. This OE-CNN, empowered by multi-feature fusion through 4-kernel parallel convolution, achieves a 95.08% MNIST classification accuracy—representing a 9.20% improvement over its single-core counterpart. The OPU demonstration achieves multi-core parallel operation and accelerated computational speed, establishing a scalable hardware foundation for optoelectronic many-core intelligent computing. Integrating coherent interference and wavelength-division multiplexing, a four-core optical computing chip is demonstrated to support 124-channel parallel processing and achieve a computational speed of 65 TOPS.