ASCENT Theme 4 Liaison Meeting / Multi-Bit Content-Addressable Memories for Exact Match and Nearest Neighbor Search Operations — with a Hyperdimensional Computing Case Study

Multi-Bit Content-Addressable Memories for Exact Match and Nearest Neighbor Search Operations — with a Hyperdimensional Computing Case Study

Presenters:

Arman Kazemi, Mehdi Sharifi (University of Notre Dame (Ph.D. students from Michael Niemier’s group, ASCENT)
Mohsen Imani (Formerly of UC-San Diego/CRISP center, now at UC Irvine)

Abstract:
Content-addressable memories (CAMs) offer high efficiencies for nearest neighbor (NN) and exact match searches. Multi-bit CAMs (MCAMs) offer further improvements over TCAMs due to higher density, but realizing multi-bit designs is challenging. Furthermore, current MCAMs do not offer NN search support. We have developed MCAM designs based on FeFETs and RRAM that build on a previous TCAM design, and that can be realized by a tunable input/programming scheme to accommodate applications with different requirements, e.g., number of bits and noise resistance. The MCAM design is augmented with the notion of a "sampling window” (SW) to further support exact match search operations. The proposed MCAM can natively evaluate a novel sigmoid-like distance function to accomplish NN search. The proposed 3-bit MCAM supports a single-step NN search and can be used to achieve 98.34% accuracy for a 5-way 5-shot classification task with a memory-augmented neural network for the Omniglot dataset (only 0.8% lower than software-based implementations). This represents a 23.5% improvement over state-of-the-art TCAM-based implementations with the same delay, energy, and area consumption. In conjunction with CRISP center personnel, we have evaluated MCAM-based distance measurements in the context of hyperdimensional computing (HDC) applications. We observe that MCAM not only improves HDC computation efficiency but also significantly boosts HDC accuracy by using a more precise distance metric.  For an example of classification, MCAM improves the HDC accuracy by 10% and replaces all costly associative search in HDC with an efficient in-memory search. 

Task: 2776.042

This meeting is only available to the JUMP research community, such as Principal Investigators, Postdoc researchers, Students, and Industry/Government liaisons.