RETROSPECTIVE:PRIME: A Novel Processing-in-memory Architecture for Neural Network Computation in ReRAM-based Main Memory

Abstract

This paper was originally published in ISCA 2016. One notable event during that ISCA was the inauguration of the Architecture 2030 Workshop [1], led by visionary leaders within our community. This workshop provided a platform for envisioning the trajectory of our field and discussing future challenges and opportunities. During this workshop, I had the privilege of presenting my vision on ”Technology-driven Architecture Innovation: Challenges and Opportunities.” To exemplify the potential for novel architecture designs, I highlighted this ISCA 2016 paper, showcasing the possibilities that arise from combining emerging applications such as AI with cutting-edge technologies like ReRAM. The workshop resulted in a report that outlines the collective vision of computer architecture research for the next 15 years, highlighting five high-potential directions: specialized accelerator design, cloud computing, 3D stacked architecture, technology-driven architecture (e.g., new memory, quantum, etc.), and machine learning as a key workload. Remarkably, our ISCA 2016 paper serves as an excellent illustrative example of three out of these five directions: emerging memory technology, domain-specific accelerators, and machine learning workloads.