Hop-CIM: An All-Digital Two-level Approximate SRAM-CIM Macro for High Energy-Efficient HNN Acceleration with Data-Aware Early Exit Mechanism

Hopfield Neural Networks (HNNs) have demonstrated effectiveness in image restoration. However, due to the intrinsically error tolerance arising from the state update criterion, performing the full precision Vector-Matrix Multiplication (VMM) to update the HNNs’ states results in computational redundancy. To address this challenge, Hop-CIM, an all-digital, two-level approximate SRAM-CIM macro is proposed in this paper with a data-aware early exit mechanism to achieve higher energy efficiency. Experimental results show that under the 28nm technology node, the proposed Hop-CIM demonstrates minimal corruption when performing image restoration on 28×28 images, as compared to the ground truths. Meanwhile, it achieves an overall energy efficiency of 941 TOPS/W, with an improvement of 1.47x and 4.44x over the traditional approximate and full precision all-digital SRAM-CIM based designs, respectively.