Crossbar Signal Shuffling for Defect-Tolerant Computing Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Crossbar arrays in computing environments suffer from defects such as un-tunable cross-point devices, leading to computing inaccuracy and high costs due to the need for redundancy or re-training neural networks, which are impractical and costly.
Innovation Solution
A system and method that includes a shuffling module to divert input signals to alternative rows within a crossbar array based on defect patterns, using a crossbar-based switch array and compensation modules to produce corrected output signals, thereby mitigating defects without degrading performance or increasing costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If defect mitigation techniques are implemented in crossbar arrays, then computing accuracy is improved, but device complexity increases
Solution Approach 1:
The system performs preliminary defect detection and characterization before computing operations. The defect pattern is identified and stored in advance, allowing the shuffling module to pre-determine optimal signal routing paths that avoid defective devices, thereby improving computing accuracy without adding complex real-time correction mechanisms during operation.
Solution Approach 2:
A shuffling module is introduced as an intermediary component between the input signal source and the crossbar array. This module dynamically routes signals through alternative paths based on pre-characterized defect patterns, effectively mediating the interaction between signals and defective devices to maintain computing accuracy while avoiding direct engagement with problematic elements.
2Measurement precision
If redundant rows are added to compensate for defective devices, then computing accuracy is improved, but manufacturing cost increases
Solution Approach 1:
Instead of adding physical redundant rows to the crossbar array, the system changes the routing parameters of existing signal paths. The shuffling module dynamically adjusts which row receives the input signal based on pre-characterized defect patterns, effectively utilizing existing hardware resources through parameter optimization rather than physical expansion.
Solution Approach 2:
The system creates a virtual copy of the crossbar array's functional capability through software-based shuffling logic rather than physical duplication. The defect pattern information is stored and used to generate alternative routing configurations, effectively copying the computational function through information processing rather than hardware replication.
3Ease of manufacture
If the crossbar array uses defective devices with un-tunable conductance, then manufacturing cost is reduced, but computing accuracy deteriorates
Solution Approach 1:
The system accepts defective devices with un-tunable conductance as inevitable manufacturing outcomes but converts this harm into a benefit by pre-characterizing their defect patterns. The shuffling module then uses this information to strategically route signals away from defective devices, transforming the presence of defects into an opportunity for optimized signal routing that maintains or even improves computing accuracy while accepting lower manufacturing costs.
Data Source
AI summary
Systems and methods for mitigating defects in a crossbar-based computing environment are disclosed. In some implementations, an apparatus comprises: a first row wire; a second row wire having a defective device; and a crossbar array configured to receive a first input signal. The defective device is associated with a defect pattern; the first row wire and the second row wire are located within the crossbar array; and the first input signal is configured to be provided to the first row wire. The apparatus further comprises a shuffling module connected to the crossbar array and configured to divert the input signal from the first row wire to the second row wire in accordance with a determination that the defect pattern is within a predefined approximation to a target circuit pattern associated with the first input signal.