(An) energy-efficient absolute-difference-accumulation operation computing-in-memory processor with computation reuse

Abstract

Recently, Computing-In-Memory (CIM) processors have been proposed to achieve high energy-efficiency by reducing data movement and solving memory bottlenecks. Furthermore, a network with highly accurate image classification has been introduced through the Absolute-Difference-Accumulation (ADA) operation instead of the multiplication-and-accumulation operation, which is widely used in DNN. ADA operation provides not only opportunities for high energy-efficient DNN accelerating by reducing multiplication but also a chance to reuse computation results. However, the previous CIM processor cannot reuse previous computation results for other computations. In this brief, we propose a highly accurate and high energy-efficient ADA-CIM processor that with two key features: 1) computation reuse for low-power, resulting in a 49.5% CIM operation power reduction, and 2) low-cost sign prediction core with 3-bit activation and weight quantization for high utilization. From the two key features, the proposed ADA-CIM processor is simulated in 28 nm CMOS technology and occupies 3.78 mm2. It consumes 2.77mW and achieves 43.1 TOPS/W energy-efficiency with a high-accuracy of 91.62% at CIFAR-10 (ResNet-20).