A Hybrid Optical-Electrical Analog Deep Learning Accelerator Using Incoherent Optical Signals

Optical deep learning (DL) accelerators have attracted significant interests due to their latency and power advantages. In this article, we focus on incoherent optical designs. A challenge is that there no known solution perform single-wavelength accumulation (a key operation required for DL workloads) using signals efficiently. Therefore, devise a hybrid approach, where done in the electrical domain, multiplication performed domain. The technology enabler of our design transistor laser, which performs electrical-to-optical optical-to-electrical conversions Through detailed evaluation design, along with comprehensive benchmarking study against state-of-the-art RRAM-based designs, derive following results: (1) For four-layer multilayer perceptron network, achieves 115× 17.11× improvements energy, respectively, compared design. We can take full advantage speed energy benefits because inference task be entirely mapped onto (2) complex workload (Resnet50), weight reprogramming needed, intermediate results need stored/re-fetched to/from memories. case, same area, still outperforms by 15.92× latency, 8.99× energy.