42 TOPS/mm2 Photonic Convolutional Processor Empowered by a Soliton Comb and a Microdisk Resonator Array

Photonic convolutional accelerators can effectively mitigate the speed bottleneck of traditional electronic processors by harnessing the intrinsic high speed and parallelism of photonics, significantly boosting the computing speed while reducing energy consumption. However, the degraded wavelength stability and poor miniaturization of bulky free-running light sources, together with the limited scalability of chip-scale convolutional cores, constrain the computing density and operational stability of conventional photonic processors. To address these challenges, we propose and demonstrate a photonic convolutional processor incorporating a compact soliton microcomb and a silicon integrated field programmable disk mesh (FPDM). The microcomb integrates a high-Q multimode Fabry-Perot resonator and a pump laser diode with self-injection locking for stabilization, enabling the generation of an ultra-stable soliton comb with a low phase noise of -126 dBc/Hz@10 kHz offset frequency and a small intensity fluctuation of less than 0.44% per hour. The FPDA has thirty-two tunable microdisks and sixteen optical I/O ports, supporting two 4-by-4 or eight 2-by-2 programmable convolution kernels for high-speed and parallel convolution processing. Moreover, featuring high scalability, reconfigurability, and support for signed operations, our photonic convolutional core achieves a record-breaking computing density of 42 TOPS/mm² and can theoretically scale up to 84 TOPS/mm². These advancements substantially enhance the scalability and adaptability of optical deep learning accelerators, paving the way for next-generation fully integrated photonic AI hardware.