Hybrid Neural Memory Array for Analog-Digital Weight Programming
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current artificial neural networks face challenges in high-performance information processing due to inadequate hardware technology, particularly in achieving high connectivity and energy efficiency, with existing digital solutions being costly and energy-inefficient compared to biological networks.
Innovation Solution
A hybrid memory system that can store weight data in both analog and digital forms using non-volatile memory arrays, allowing for continuous programming and reading of memory cells without disturbing other cells, enabling precise tuning of synapse weights in neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If digital solutions are used for artificial neural networks, then computational precision is improved, but energy efficiency deteriorates and cost increases
Solution Approach 1:
The memory system is designed to perform multiple functions: it can operate as a digital memory system for precise data storage and as an analog neural network system for energy-efficient computation. The same physical memory array supports both digital read/program operations and analog neural network inference operations, eliminating the need for separate hardware systems.
Solution Approach 2:
The patent replaces traditional digital CMOS logic circuits (which perform multiplication and addition separately) with an analog memory-based system that performs these operations simultaneously through physical phenomena. The memory array uses voltage division and current flow to naturally perform matrix-vector multiplication, substituting complex digital logic with simpler analog physical processes.
2Measurement precision
If separate multiplication and addition logic circuits are used, then computational accuracy is improved, but device complexity increases
Solution Approach 1:
The patent merges multiplication and addition operations into a single simultaneous process. The memory array performs both operations at the same time through analog current summation, eliminating the need for separate logic circuits. This combining of operations reduces device complexity while maintaining computational accuracy through the inherent linearity of analog circuits.
3Use of energy by moving object
If analog computation is used to improve energy efficiency, then energy consumption is reduced, but manufacturing precision requirements increase
Solution Approach 1:
The memory system performs self-calibration and self-adjustment to compensate for manufacturing variations. The system includes mechanisms for measuring and adjusting threshold voltages and weight values to account for process variations, allowing the analog neural network to achieve accurate computation despite inherent manufacturing imprecision in the memory cells.
Data Source
AI summary
Numerous embodiments of a hybrid memory system are disclosed. The hybrid memory can store weight data in an array in analog form when used in an analog neural memory system or in digital form when used in a digital neural memory system. Input circuitry and output circuitry are capable of supporting both forms of weight data.