Microrings as programmable temporal kernels enabling photonic AI beyond 100 Gbaud
TL;DR
This work redefine MRRs as programmable temporal convolution kernels by exploiting their impulse responses, enabling computation beyond the resonance linewidth, and realizes a multi-channel MRR convolution engine with 120.8 TOPS throughput and reduce the hardware complexity.
Abstract
The rapid growth of artificial intelligence (AI) demands high-performance hardware accelerators. Photonic computing with microring resonators (MRRs) has attracted significant interest, but conventional architectures use MRRs primarily as scalar weights, with speed constrained by the resonance linewidth of ~ 10 GHz. Here, we redefine MRRs as programmable temporal convolution kernels by exploiting their impulse responses, enabling computation beyond the resonance linewidth. A single MRR operating at a symbol rate of 128 Gbaud achieves a computing throughput of 5.12 trillion operations per second (TOPS), representing a 160-fold improvement over scalar weighting. By integrating temporal convolution with wavelength- and space-division multiplexing, we realize a multi-channel MRR convolution engine with 120.8 TOPS throughput and reduce the hardware complexity from O( N 2 ) to O( N ), achieving a compute density of 48.05 TOPS/mm 2 . The framework is experimentally validated through optical modulation format identification, network anomaly detection, and image classification, establishing a scalable computing primitive for photonic AI accelerators.
