Dynamic Voltage-Conductance Point Distribution for ANN Memory Arrays
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for updating the distribution of voltage-conductance points in Artificial Neural Networks (ANNs) are prone to errors due to noise sources in the analog domain, leading to inaccuracies in storing and reading values from multi-level cells, especially when uniform voltage levels are assumed for conductance levels, resulting in non-discernable voltage-conductance points that cause errors in inference generation.
Innovation Solution
The solution involves updating the distribution of voltage-conductance points based on operational noise sources and their modeling, using a quantization process that dynamically adjusts conductance levels to accurately program and read parameters in memory cells, thereby improving inference accuracy without re-training the ANN, by identifying discernable conductance levels for parameters with the most information and mapping them to distinct voltage-conductance points.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If uniform voltage levels are assumed for conductance levels in multi-level cells, then the quantization process is simplified, but errors occur in storing and reading values due to non-discernable voltage-conductance points
Solution Approach 1:
The patent dynamically adjusts voltage levels based on operational noise sources and characteristics. Instead of using fixed uniform voltage levels, the system modifies voltage parameters adaptively to account for noise variations, ensuring discernable voltage-conductance points while maintaining quantization functionality
Solution Approach 2:
The system incorporates feedback mechanisms that monitor operational noise sources and adjust voltage-conductance points accordingly. By continuously adapting voltage levels based on observed noise characteristics, the system maintains accurate parameter storage and retrieval without requiring re-training of the ANN
2Reliability
If voltage-conductance points are updated dynamically based on noise sources, then inference accuracy is improved, but the complexity of the quantization process increases
Solution Approach 1:
The system performs self-adjustment of voltage-conductance points by automatically detecting and adapting to operational noise sources. This self-service mechanism updates the quantization parameters without external intervention or re-training, maintaining inference accuracy while managing complexity through autonomous operation
Solution Approach 2:
The patent transitions from static uniform voltage levels to dynamic adaptive voltage-conductance points. The system continuously adjusts voltage levels based on operational conditions, making the quantization process flexible and responsive to changing noise characteristics, thereby improving reliability
3Measurement precision
If discernable conductance levels are identified for parameters with the most information, then parameter reading accuracy is improved, but the time required for quantization increases
Solution Approach 1:
The patent applies different voltage-level strategies to different parameters based on their information content. Parameters with the most information are assigned to discernable conductance levels with higher precision, while less critical parameters use standard quantization, optimizing overall accuracy without uniformly increasing quantization time for all parameters
Data Source
AI summary
Updating a distribution of voltage-conductance points for artificial neural networks can include receiving, at an accelerator, such as a MAC unit, information corresponding to a memory array of the MAC unit. A plurality of parameters of the ANN can be received. The distribution of voltage-conductance points can be identified utilizing the plurality of parameters and based on the information corresponding to the memory array. The distribution of voltage-conductance points correspond to discernable conductance levels and a subset of the plurality of parameters. The ANN can be stored in the MAC unit based on the discernable conductance levels.