ABSTRACT In photonic computing architectures for Convolutional Neural Networks (CNN), the Time‐Wavelength Interleaved (TWI) scheme is a key strategy for achieving high computational throughput, as it decouples computational parallelism from the number of modulators. However, this scheme has long been hindered by a fundamental mismatch between algorithmic data patterns and hardware execution constraints—a bottleneck known as “sliding‐window redundancy.” Under traditional geometry‐driven scheduling, this bottleneck limits hardware utilization to for a convolution kernel, severely restricting the realization of its theoretical computational power and energy efficiency advantages. To address this, we propose a physics‐guided scheduling theory—Shift‐Equivariant Kernel Scheduling (SEKS). This method fundamentally eliminates such redundancy by leveraging the inherent shift‐equivariance of convolution operations, achieving near 100% channel utilization. Validation based on a monolithically integrated thin‐film lithium niobate chip demonstrates that SEKS delivers a computational throughput of 1.056 TOPS for 44 convolution and achieves 97.95% accuracy in a handwritten digit recognition task. This work provides a complete co‐design framework for the TWI architecture, spanning algorithm, scheduling, and hardware implementation, thereby clearing a key obstacle on its path to practical application.
Wang et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: