In-Memory Nearest Neighbor Search with Content-Addressable Memories
Presenter: Arman Kazemi (Prof. Michael Niemier's Group, Notre Dame)
2776.042: Content Addressable Memories for Neuro-Inspired Computing
2776.083: Application-level Benefits of Emerging, Embedded, Non-volatile Memories
Abstract: Nearest neighbor (NN) search computations are at the core of many applications such as few-shot learning, kNN classification, hyperdimensional computing, etc. NN search using traditional von Neumann architectures can be quite expensive due to the memory wall, especially for high dimensional data, and efficient hardware support is desirable. In-memory computing architectures -- especially those based on emerging devices -- offer attractive solutions for NN search. Ternary content-addressable memories (TCAMs) in particular have been shown to offer substantial energy and latency improvements for NN search. However, the binary precision of TCAMs can lead to severe degradation in application-level accuracies or high memory demand, thus diminishing gains with respect to other figures of merit (FOM). Recently analog and multi-bit content-addressable memories (ACAMs and MCAMs) were proposed that offer further improvements in density and energy. However, ACAMs and MCAMs have not been explored for NN search applications.
In this presentation, we consider MCAMs for NN search with Euclidean-squared and sigmoid-like distance functions -- primarily using ferroelectric field-effect transistors (FeFETs) as context, but will show extensibility to other devices as well. MCAMs can achieve accuracies comparable to software implementations for few-shot learning applications -- even with subsets of the ImageNet dataset. We further consider the effects of device-to-device variations on FOM such as application-level accuracy and area. Notably, MCAMs achieve 2.5x higher area and energy efficiencies than TCAMs for similar accuracies. We conclude with a discussion of (i) benefits of using MCAMs for hyperdimensional computing applications, (ii) the applicability of the MCAM designs to other technologies such as RRAM and Flash, and (iii) experimental demonstrations of a 2-bit implementation of FeFET MCAMs using AND arrays from GLOBALFOUNDRIES to further validate the design concept.
Bio: Arman Kazemi is a Ph. D. student at the University of Notre Dame in the hardware-software codesign lab. His research interests include hardware/software co-design, low-power hardware design, and in-memory computing. He is particularly interested in inventions leveraging emerging beyond-CMOS technologies e.g. ferroelectric materials. His research usually targets reducing computational resource requirements of machine learning applications using emerging circuits and architectures.
This meeting is only available to the JUMP research community, such as Principal Investigators, Postdoc researchers, Students, and Industry/Government liaisons