The Road to Widely Deploying Processing-in-Memory: Challenges and Opportunities

• Published in: Proceedings / IEEE Computer Society Annual Symposium on VLSI pp. 259 - 260
• Main Author: Ghose, Saugata
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.07.2022
• Subjects: Computer architecture; hardware/software co design; Machine learning; Pressing; processing in memory; processing near memory; processing using memory; Prototypes; Roads; system integration; Technological innovation; Very large scale integration
• DOI: 10.1109/ISVLSI54635.2022.00057

Processing-in-memory (PIM) refers to a computing paradigm where some or all of the computation for an ap-plication is moved closer to where the data resides (e.g., in main memory). While PIM has been the subject of ongoing research since the 1970s [8], [11], [17], [19], [26], 2[8], [29], [33], it has experienced a resurgence in the last decade due to (1) the pressing need to reduce the energy and latency overheads associated with data movement between the CPU and memory in conventional systems [6], [18], and (2) recent innovations in memory technologies that can enable PIM integration (e.g., [13]-[16], [20], [21], [24], [31]). Recently-released products and prototypes, ranging from programmable near-memory pro-cessing units [7], [36] to custom near-bank accelerators for machine learning [22], [23], [30] and analog compute support within memory arrays [9], [27], have demonstrated the viability of manufacturing PIM architectures.