NVM Crossbar Correction Columns for ANN Noise Suppression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Non-volatile memory (NVM) accelerators with analog crossbars generate undesirable noise due to spurious currents from NVM cells programmed to a conductance value of goff, which reduces the accuracy of artificial neural networks (ANNs) in mobile and power-constrained devices.
Innovation Solution
An NVM crossbar with unipolar weights is introduced, featuring additional correction cells and signal lines to mitigate noise by correcting undesirable currents generated by cells programmed to goff, thereby enhancing the accuracy of ANN operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If NVM cells are programmed to conductance value goff to implement ANN weights, then the accelerator can perform GEMM operations, but spurious currents are generated causing noise that reduces ANN accuracy
Solution Approach 1:
The patent converts the harmful spurious currents generated by NVM cells programmed to goff into a beneficial correction signal. By intentionally generating these currents in a dedicated correction column and then subtracting them from the MAC column outputs, the system eliminates the noise they would otherwise introduce, thereby improving ANN accuracy while maintaining GEMM operation capability
Solution Approach 2:
The patent introduces a correction column as an intermediary component that handles the spurious currents separately. This correction column acts as a mediator that captures, processes, and eliminates the harmful currents before they can degrade the accuracy of the main computation, allowing the system to maintain both computational capability and precision
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The solution significantly reduces spurious noise contributions, improving the accuracy of ANN operations in NVM accelerators, particularly in power-constrained devices by effectively managing currents and maintaining performance.
Implementation Method 1
Each MAC cell includes one or more NVM elements programmed to a conductance level representing one of a plurality of unipolar weights of an ANN
Data Source
AI summary
A non-volatile memory (NVM) crossbar for an artificial neural network (ANN) accelerator is provided. The NVM crossbar includes row signal lines configured to receive input analog voltage signals, multiply-and-accumulate (MAC) column signal lines, a correction column signal line, a MAC cell disposed at each row signal line and MAC column signal line intersection, and a correction cell disposed at each row signal line and correction column signal line intersection. Each MAC cell includes one or more programmable NVM elements programmed to an ANN unipolar weight, and each correction cell includes one or more programmable NVM elements. Each MAC column signal line generates a MAC signal based on the input analog voltage signals and the respective MAC cells, and the correction column signal line generates a correction signal based on the input analog voltage signals and the correction cells. Each MAC signal is corrected based on the correction signal.