Work-in-Progress: A Scalable Stochastic Number Generator for Phase Change Memory Based In-Memory Stochastic Processing

Abstract

Stochastic computing based Processing-In-Memory (PIM) architectures (e.g., [1]) can provide massive parallelism with higher energy-efficiency, for implementing complex computations in main memory. However, stochastic computing arithmetic requires random bit streams generated by stochastic number generators (SNGs), which account for significant area and energy consumption. Moreover, SNGs' numerical precision needs improvement to reduce errors in stochastic computations [1]. Thus, low numerical precision and high implementation overheads of SNGs can offset the benefits of adopting stochastic computing in PIM architectures. In this paper, we exploit the inherent stochasticity of Phase Change Memory (PCM) cells to design a scalable and area-energy efficient SNG for PCM-based stochastic PIM architectures. Our designed SNG can achieve up to ~300× lower area and up to ~250× lower energy consumption with better numerical precision, compared to the Linear Feedback Shift Register (LFSR) based conventional SNG from [2].