Analog Crossbar Nonlinearity Correction for Matrix Multiply Outputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Analog crossbar matrix multiply engines introduce non-linear responses due to hardware limitations, making it challenging to accurately decouple linear and non-linear signals, which affects processing efficiency and accuracy.
Innovation Solution
A system comprising a processor and memory that determines a non-linear correction term for outputs of an array of analog memories based on matrix vector multiplication results, allowing for automatic decoupling and correction of non-linear responses, thereby improving processing performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If analog crossbar matrix multiply engines are used for high-speed computing, then processing speed is improved, but non-linear responses are introduced due to hardware limitations
Solution Approach 1:
The patent introduces a digital correction system as an intermediary between the analog computation engine and the final output. The correction system includes a correction term generator that creates non-linear correction terms and a correction application module that applies these terms to the analog outputs, thereby mediating the transition from imprecise analog results to corrected digital outputs while preserving the speed advantage of analog computing
Solution Approach 2:
The patent replaces the need for perfectly linear analog hardware with a hybrid system that uses digital signal processing to correct analog outputs. Instead of relying solely on mechanical/analog linearity, the system substitutes digital correction algorithms that can precisely compensate for non-linearities, effectively replacing the requirement for perfect analog hardware with computational correction
2Measurement precision
If non-linear correction terms are calculated and applied to all outputs, then output accuracy is improved, but computational overhead increases
Solution Approach 1:
The patent performs preliminary characterization of the analog hardware's non-linear behavior during a calibration phase. The system pre-calculates correction terms based on known input-output relationships and stores these correction characteristics for rapid application during actual computation, avoiding the need to perform complex non-linear analysis in real-time
Solution Approach 2:
The patent transforms the correction problem from calculating complex non-linear corrections for each output to applying pre-determined correction parameters. By changing the approach from real-time non-linear analysis to parameter-based correction, the system reduces computational overhead while maintaining accuracy
Data Source
AI summary
Techniques facilitating automatic non-linearity correction for analog hardware are provided. A system can comprise a memory that stores computer executable components and a processor that executes the computer executable components stored in the memory. The computer executable components can comprise an adjustment component that determines a non-linear correction term for an output of an array of analog memories based on a result of a matrix vector multiplication performed on the array of analog memories. The computer executable components can also comprise a rectification component that applies the non-linear correction term to additional outputs of the array of analog memories.