Defects Mitigation in Resistive Crossbars for Analog Vector Matrix Multiplication

TL;DR

This paper proposes defect mitigation techniques for resistive crossbars used in analog vector matrix multiplication, enabling improved accuracy without re-training or redundancy, and demonstrating effectiveness in practical applications.

Contribution

Introduction of row shuffling and output compensation methods to mitigate defects without re-training or redundant crossbars, along with analysis of defect and parasitic effects.

Findings

Rescued up to 10% of defects in ResNet-20 without performance loss.

Analyzed coupling effects of defects and circuit parasitics.

Achieved a good trade-off between cost and performance.

Abstract

With storage and computation happening at the same place, computing in resistive crossbars minimizes data movement and avoids the memory bottleneck issue. It leads to ultra-high energy efficiency for data-intensive applications. However, defects in crossbars severely affect computing accuracy. Existing solutions, including re-training with defects and redundant designs, but they have limitations in practical implementations. In this work, we introduce row shuffling and output compensation to mitigate defects without re-training or redundant resistive crossbars. We also analyzed the coupling effects of defects and circuit parasitics. Moreover, We study different combinations of methods to achieve the best trade-off between cost and performance. Our proposed methods could rescue up to 10% of defects in ResNet-20 application without performance degradation.