Memristor Crossbar Array Weight Update Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Updating model/weight data in a memristor crossbar array during the training phase of deep neural networks can lead to processing inefficiencies.
Innovation Solution
A memristive dot product system that includes a crossbar array with memristor and transistor elements, a comparator processor, and analog-to-digital and digital-to-analog converters to efficiently compare and update input vector data with existing data in the memristor array, reducing unnecessary write operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If model/weight data is continuously updated in the memristor crossbar array during training phase, then the model can be trained with current data, but processing efficiency decreases due to unnecessary write operations
Solution Approach 1:
The system performs preliminary comparison between new weight values and existing memristor states before executing write operations. The comparator processor evaluates whether weight changes exceed a threshold, and only triggers updates when necessary, preventing premature or redundant write operations to the crossbar array.
Solution Approach 2:
The system implements a feedback mechanism where the comparator processor continuously monitors weight changes and provides control signals to the write operation controller. This feedback loop ensures that updates are only performed when actual changes occur, creating a closed-loop control system that optimizes write operations based on real-time data.
2Manufacturing precision
If all weight changes are written to the memristor crossbar array, then the model training is complete and accurate, but processing time increases
Solution Approach 1:
The system extracts only the essential weight changes that exceed a predefined threshold from the complete set of weight updates. By separating significant changes from negligible ones, the system performs only necessary write operations to the crossbar array, reducing processing time while maintaining training accuracy.
Solution Approach 2:
The system changes the parameter of weight update threshold dynamically or statically to control when write operations occur. By adjusting this parameter, the system balances between updating all weights (higher accuracy) and updating only significant changes (lower processing time), optimizing the trade-off based on training requirements.
3Reliability
If the memristor crossbar array is iteratively updated during training, then the model converges to optimum weights, but energy consumption increases due to frequent write operations
Solution Approach 1:
The system performs preliminary evaluation of weight changes before committing them to the crossbar array. By pre-assessing whether changes warrant an update, the system avoids energy-intensive write operations for negligible changes, reducing overall energy consumption during the iterative training process.
Solution Approach 2:
The feedback mechanism monitors weight change magnitudes and controls write operation activation accordingly. This feedback loop creates an energy-efficient training process by enabling writes only when necessary, reducing the frequency of high-energy operations while maintaining model convergence.
Data Source
AI summary
A crossbar array includes a number of memory elements. An analog-to-digital converter (ADC) is electronically coupled to the vector output register. A digital-to-analog converter (DAC) is electronically coupled to the vector input register. A processor is electronically coupled to the ADC and to the DAC. The processor may be configured to determine whether division of input vector data by output vector data from the crossbar array is within a threshold value, and if not within the threshold value, determine changed data values as between the output vector data and the input vector data, and write the changed data values to the memory elements of the crossbar array.