A Hybrid Opto-Electrical Floating-point Multiplier

The performance improvement by CMOS circuit technology is reaching its limits. Many researchers have been studying computing technologies that use emerging devices to challenge such critical issues. Nanophotonic technology is a promising candidate due to its ultra-low latency, high bandwidth, and low power natures. The advanced research activity of nanophotonic computing is to design hardware accelerators for AI inference applications. However, few considerations about nanophotonic accelerators for AI training applications have been conducted. The main reason is that state-of-the-art nanophotonic AI accelerators involve integer operations, whereas floating-point (FP) sum-of-products dominate the training process. However, to the best of the authors' knowledge, there are no optical circuits that target floating-point arithmetic units. This study proposes a novel Opto-Electrical Floating-point Multiplier (OEFM) toward ultra-low-latency, a power-efficient nanophotonic accelerator for AI training applications. We design a microarchitecture of OEFM, including a novel optical integer multiplier and other electrical components. Based on our evaluation framework, we analyze the calculation accuracy of the proposed multiplier and OEFM. Experimental results show that OEFM achieves a 56 % reduction in latency and a 41 % reduction in energy consumption compared with a conventional electrical circuit.