DIET-PIM: Dynamic Importance-based Early Termination for Energy-Efficient Processing-in-Memory Accelerator

Processing-In-Memory (PIM) has notably boosted the energy efficiency of Convolutional Neural Networks (CNN) inference. However, frequent access to power-hungry analog-to-digital converters (ADCs) remains challenging, hindering overall energy efficiency. Although ongoing research focuses on exploiting sparsity to address this problem, the ADC overhead has not been sufficiently reduced due to static inference. This paper proposes a dynamic importance-based early termination technique for PIM (DIET-PIM). Specifically, DIET-PIM optimizes CNN inference at runtime by identifying regions in the features that contribute less to the classification result and skipping the computation on a subset of the input values for these ineffective regions. Evaluation results show that DIET-PIM reduces ADC accesses and improves overall energy efficiency by $3.13 \times-5.79 \times$ compared with baseline PIM-based accelerator.