[C113] Partial-Sum Regularization for Energy-Efficient ADC-Less CIM Accelerators

Abstract

In this work, we propose a partial-sum (PS) regularization method for the energy-efficient ADC-less compute-in-memory (CIM) accelerator. In this accelerator, ADCs are replaced by comparators that quantize PSs into ternary values, and the corresponding scale-factor computations are performed in digital CIM (DCIM). Since DCIM operations can be skipped for zero-valued PSs, the DCIM energy consumption strongly depends on the number of non-zero PSs. However, existing ternary PS quantization does not explicitly optimize PS sparsity during training; instead, PS sparsity emerges only as a by-product of accuracy-driven training. To address this issue, the proposed method directly penalizes non-zero ternary PSs generated under the target CIM array mapping. In addition, we introduce bit significance weighting (BSW) that imposes stronger penalties on PSs from less significant weight bits to increase PS sparsity while mitigating accuracy degradation. Evaluations on ResNet-20/CIFAR-10 and ResNet-18/ImageNet show that the proposed method improves PS sparsity by up to 12.7%p, reduces per-bit DCIM energy by up to 30%, and improves system-level energy efficiency by up to 8.9%.

Publication
Design, Automation, and Test in Europe 2027
Jong Hwan Ko (고종환)
Jong Hwan Ko (고종환)
Associate Professor