STAR: A Mixed Analog Stochastic In-DRAM Convolutional Neural Network Accelerator

Data movement remains the biggest energy and performance bottleneck for accelerating machine learning workloads in modern computing platforms. We introduce STAR, a mixed analog/stochastic-based in-DRAM accelerator for convolutional neural network (CNN) inference tasks. We propose lightweight modifications and circuits into conventional DRAM arrays, to support stochastic computing for multiplications and devise a novel in-DRAM metal-on-metal capacitor for analog accumulations. STAR combines in-sense-amplifier and near-subarray computing to overcome challenges inherent in prior in-DRAM and stochastic-based accelerators. Experimental results indicate that STAR exhibits at least 2.1× lower latency, 2.6× better computational efficiency, and 30% higher power efficiency compared to state-of-the-art in-DRAM accelerators.