PIM-DRAM: Accelerating Machine Learning Workloads Using Processing in Commodity DRAM

Deep Neural Networks (DNNs) have transformed the field of machine learning (ML) and are widely deployed in many applications involving image, video, speech natural language processing. The increasing compute demands DNNs been addressed through Graphics Processing Units (GPUs) specialized accelerators. However, as model sizes grow, these von Neumann architectures require very high off-chip memory bandwidth to keep processing elements utilized, a majority data resides main memory. is actively being explored promising solution wall bottleneck for ML workloads. In this work, we propose new DRAM-based processing-in-memory (PIM) multiplication primitive coupled with intra-bank accumulation accelerate matrix vector multiply operations proposed adds <1% area overhead does not any change DRAM peripherals. Subsequently, design PIM architecture (PIM-DRAM) mapping scheme executing on architecture. System evaluations performed AlexNet, VGG16 ResNet18 show that architecture, mapping, flow can provide up 19.5x speedup over an NVIDIA Titan Xp GPU, highlighting potential future generations DNN hardware.