Software-controlled high-dimensional coherent optical tensor computing via wavelength multiplexing

With the growing demand for real-time data processing in artificial intelligence applications, optical computing has emerged as a promising solution because of its inherent high speed and parallelism. In particular, high-dimensional optical tensor computing offers an efficient solution for high-throughput inference in optoelectronic convolutional neural networks (OE-CNNs). Here, we present a software-controlled reconfigurable optical tensor processing unit that leverages both optical coherent technology and wavelength-division multiplexing (WDM) to enable high-dimensional matrix operations and high-throughput computing. The chip integrates nine Mach-Zehnder interferometers (MZIs) along with nine matched optical delay lines to perform coherent serial 3 × 3 convolution operations with software-tunable ternary weights of {-1, 0, 1}. WDM is further introduced to support high-dimensional tensor computing, allowing parallel processing of multiple data channels. The chip experimentally executes tensor computing on 3-dimensional images at a data rate of 29.14 GBaud, achieving a peak computational speed of 1.57 TOPS. With further expansion to 37-channel WDM, it can achieve a potential computational speed of up to 19.41 TOPS. An OE-CNN is further demonstrated, comprising four parallel optical convolution kernels within a single convolutional layer, followed by an electrical fully connected layer. Moreover, the OE-CNN achieves a classification accuracy of 96.66% on the MNIST dataset. By seamlessly integrating the spectral efficiency of coherent optical computing with the high parallelism of WDM, the proposed architecture significantly enhances the photonic computing performance. This work provides a scalable pathway toward the development of next-generation high-performance photonic AI processors.