Compute-in-Memory와 Time-Domain 연산을 활용한 Spiking Neural Network 가속기

Abstract

Spiking Neural Network (SNN) draws attention for low energy consumption due to transmitting only binary information. Another way to implement low power hardware is Compute-in-Memory (CIM). It removes external memory access and minimizes data transaction energy. However, power-hungry Analog-to-Digital Converter (ADC) is essential in analog CIM. In time-domain computation, ADC is replaced with time-to-digital converter. Nevertheless, time-domain CIM supporting multi-bit computation has low energy efficiency by digital logic. To resolve this issue, this paper proposes time-domain SNN CIM with no digital circuit and it improves 83% throughput by pipelined architecture. It achieves 701.7 TOPS/W of energy efficiency.