Abstract:
In-memory computing-based systems deliver en-hanced performance by eliminating data movement between computational and storage units. Among various in-memory computing pr...Show MoreMetadata
Abstract:
In-memory computing-based systems deliver en-hanced performance by eliminating data movement between computational and storage units. Among various in-memory computing primitives, Content Addressable Memories (CAMs) utilizing resistive memory technologies have gained attention for their highly parallel and energy-efficient pattern-matching capabilities. In this study, we explore the trade-offs and challenges involved in designing an Analog CAM array and configuring a multi-dimensional parameter space. We integrate Analog CAM arrays and peripheral circuits on-chip using the TSMC 28nm process, along with BEOL-integrated TaOx resistive memories. By mapping tree-based machine learning models onto the ACAM array, we showcase low-latency, energy-efficient, single-cycle accelerated inference performance.
Date of Conference: 06-09 August 2023
Date Added to IEEE Xplore: 31 January 2024
ISBN Information: