Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Content Addressable Memory Design with Reference Resistor for Improved Search Resolution

Siri Narla, Steven J. Koester, Rebecca A. Dawley, Ageeth A. Bol, Piyush Kumar, Azad Naeemi

TL;DR

Content-addressable memories offer fast in-memory similarity search but are limited by resolution as arrays scale. The authors introduce a BEOL-compatible fixed reference resistor in the discharge path and a prolonged precharge scheme to make matchline discharge depend mainly on the reference, enabling high-resolution search. Nb-doped WS2 resistors fabricated by ALD demonstrate near-ideal resolution of $\leq 5$-bits across SOT-R, FeFET-R, and SRAM-R CAMs, with improved variation tolerance at the cost of increased energy and delay. The approach supports scalable CAM implementations for AI workloads by delivering robust search performance on advanced nodes.

Abstract

Despite the parallel in-memory search capabilities of content addressable memories (CAMs), their use in applications is constrained by their limited resolution that worsens as they are scaled to larger arrays or advanced nodes. In this work we present experimental results for a novel back-end-of-line compatible reference resistive device that can significantly improve the search resolution of CAMs implemented with CMOS and beyond-CMOS technologies to less than or equal to 5-bits.

Content Addressable Memory Design with Reference Resistor for Improved Search Resolution

TL;DR

Content-addressable memories offer fast in-memory similarity search but are limited by resolution as arrays scale. The authors introduce a BEOL-compatible fixed reference resistor in the discharge path and a prolonged precharge scheme to make matchline discharge depend mainly on the reference, enabling high-resolution search. Nb-doped WS2 resistors fabricated by ALD demonstrate near-ideal resolution of -bits across SOT-R, FeFET-R, and SRAM-R CAMs, with improved variation tolerance at the cost of increased energy and delay. The approach supports scalable CAM implementations for AI workloads by delivering robust search performance on advanced nodes.

Abstract

Despite the parallel in-memory search capabilities of content addressable memories (CAMs), their use in applications is constrained by their limited resolution that worsens as they are scaled to larger arrays or advanced nodes. In this work we present experimental results for a novel back-end-of-line compatible reference resistive device that can significantly improve the search resolution of CAMs implemented with CMOS and beyond-CMOS technologies to less than or equal to 5-bits.
Paper Structure (8 sections, 3 figures)

This paper contains 8 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Resistive Elements
  3. Cell Design and Modeling Framework
  4. Results
  5. Resolution with Nominal Devices
  6. Variability Study
  7. Energy and Delay Results
  8. Conclusion

Figures (3)

  • Figure 1: (a) TLM data and sheet resistance extraction of ALD TMD material. Inset: cross-section transmission electron microscopy image of ALD Nb-doped WS2 (b) Resistance vs. voltage for resistors with varying contact spacings.
  • Figure 2: (a) SOT-R (b) SRAM-R and (c) FeFET-R CAM cell schematics
  • Figure 3: (MDD as a function of HDist for a 128x128 (a) SOT-R CAM array with various reference resistances, (b) SOT, FeFET and SRAM-based CAM arrays with and without modifications, (c) SOT, SOT-R and SOT-R PRE CAMs with and without variation (var), and (d) Search Energy and Delay results for HDist=20.