Tensor convolution is a fundamental operation in convolutional neural networks, especially for processing tensors, which are prevalent in real-world applications. Current methods often convert tensor convolutions into matrix multiplications, leading to data replication, additional memory usage and increased hardware complexity. Here, a high-bit-efficiency optical tensor convolution accelerator with reduced data redundancy and lower memory consumption is presented. The bit-efficiency of the optical tensor convolution accelerator is first explored, significantly improving its effective computing power by utilizing the spatial dimension. Consequently, the optical tensor convolutional accelerator operates at speeds exceeding 3 Tera Operations Per Second (TOPS)—the fastest single-kernel optical convolutional accelerator to date, to the best of authors' knowledge. Its performance is validated on handwritten digit recognition and histopathologic cancer detection tasks, achieving 93.8% and 77% accuracy, respectively, closely matching in-silico results. This approach simultaneously multiplexes the physical dimensions—wavelength, time, and space—and leverages the parallelism and high throughput of light, enabling efficient optical processing of tensor data with significant computational power.
原文链接
分享
引用
