RedPIM: An Efficient PIM Accelerator Design with Reduced Analog-to-Digital Conversions

ReRAM-based Processing-In-Memory (PIM) architectures are compelling contenders for deep learning due to their ability to perform matrix-vector multiplications (MVMs) directly within the memory, significantly reducing data movement and enhancing computational efficiency. However, since MVMs occur in the analog domain, analog-to-digital converters (ADCs) dominate power consumption and area overhead in current implementations. To this end, we propose RedPIM, an efficient ReRAM-based PIM accelerator design for deep neural networks (DNNs) that reduces the number of analog-to-digital conversions. RedPIM exploits the fact that in ReRAM-based PIM accelerators, the overall energy consumption generally increases with the number of activated analog-to-digital conversions. Specifically, we introduce a novel training algorithm that is aware of the ADC overhead during activation value quantization and optimizes accuracy concurrently. From a hardware design perspective, we develop a lookup table (LUT)-based quantization module to enable efficient and low-cost activation value quantization. In addition, we propose an efficient adaptive operation unit (OU) size assignment scheme that further minimizes analog-to-digital conversions by considering activation sparsity and weight distribution. Extensive experimental results show our RedPIM reduces latency to 27.72% and energy consumption to 10.15% of the baseline, with minimal accuracy loss, making it a promising solution for enhancing DNN acceleration. The code for this project is available at: https://github.com/JialeLiLab/ADC_aware_Learning.git.