Gain Cell-Based Analog Content Addressable Memory for Dynamic Associative tasks in AI

TL;DR

This work proposes a capacitor gain cell-based aCAM designed for dynamic processing, and demonstrates the application of aCAM within transformer attention mechanisms by replacing the softmax-scaled dot-product similarity with aCAM similarity, achieving competitive results.

Abstract

Analog Content Addressable Memories (aCAMs) have proven useful for associative in-memory computing applications like Decision Trees, Finite State Machines, and Hyper-dimensional Computing. While non-volatile implementations using FeFETs and ReRAM devices offer speed, power, and area advantages, they suffer from slow write speeds and limited write cycles, making them less suitable for computations involving fully dynamic data patterns. To address these limitations, in this work, we propose a capacitor gain cell-based aCAM designed for dynamic processing, where frequent memory updates are required. Our system compares analog input voltages to boundaries stored in capacitors, enabling efficient dynamic tasks. We demonstrate the application of aCAM within transformer attention mechanisms by replacing the softmax-scaled dot-product similarity with aCAM similarity, achieving competitive results. Circuit simulations on a TSMC 28 nm node show promising performance in terms of energy efficiency, precision, and latency, making it well-suited for fast, dynamic AI applications.

Figures (3)

• Figure 1: (a) Circuit configuration of the proposed dynamic aCAM cell. (b) Architecture of the dynamic IMC aCAM macro including peripheral circuitry. (c&d) Input voltage range within the dynamic search window for a given stored key voltage $V_{store}$ in both mismatch (c) and match (d) cases.
• Figure 2: (a) Cumulative Distribution Function (CDF) from 50 Monte Carlo simulations, showing the distribution of ML current amplitudes for match (blue) and mismatch (green) conditions in a dynamic aCAM cell. Match condition: $V_{Store} = 0.45 \, \textrm{V}$, $\textrm{VDL} = 0.45 \, \textrm{V}$; Mismatch condition: $V_{Store} = 0.45 \, \textrm{V}$, $\textrm{VDL} = 0.40 \, \textrm{V}$. (b) Confusion matrix representing the results of 50 Monte Carlo simulations for all combinations of a search input and a stored reference. Diagonal entries indicate matching cases where $V_{store} = \textrm{VDL}$. A match is determined by comparing the measured current to the threshold $I_{match} = 350 \, \textrm{nA}$. The number of matches exceeding this threshold is counted, which is indicated by the numbers in each pixel.
• Figure 3: (a) Conventional attention head with scaled dot product similarity (SDPS). (b) proposed attention head using the dynamic aCAM for computing the similarity between $K$ and $Q$. (c) Electromyographic data processing task where the transformer predicts finger-joint angles. The predicted angles (degrees of actuation, DOA) and the target angles are respectively represented by red and black curves, while the input electromyographic signal is represented in blue. (d) Training results comparing conventional attention-based transformer with dynamic aCAM attention-based transformer.