Hybrid Neural Memory Array for Analog-Digital Weight Storage
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial neural networks face challenges in achieving high-performance information processing due to a lack of adequate hardware technology, particularly in terms of energy efficiency and the high computational complexity required for large numbers of synapses, which are not efficiently addressed by digital supercomputers or specialized graphics processing units.
Innovation Solution
The use of non-volatile memory arrays as synapses in an analog neural network that utilizes a hybrid memory system configurable to store neural memory arrays as synapses in a hybrid memory system configurable to store neural memory arrays configurable to store neural memory weight data in analog or digital form.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If digital supercomputers or specialized graphics processing units are used to achieve high computational parallelism for neural networks, then the computational capability is improved, but the energy efficiency deteriorates
Solution Approach 1:
The patent replaces digital computational systems (supercomputers, GPUs) with an analog neural network system using non-volatile memory arrays. The memory array directly performs analog matrix multiplication operations, substituting the mechanical/digital computation process with an analog physical process that naturally computes neural network operations through voltage distributions across the memory array, achieving both high computational capability and energy efficiency
Solution Approach 2:
The patent changes the operational parameters from digital binary states to analog continuous voltage levels. By programming non-volatile memory cells to represent analog weight values through continuous resistance changes, the system enables analog computation that mimics biological neural networks, simultaneously improving computational capability for neural networks and reducing energy consumption compared to digital systems
2Use of energy by moving object
If CMOS analog circuits are used to implement synapses in neural networks, then energy efficiency is improved, but the device area increases due to bulky circuit implementation
Solution Approach 1:
The patent extracts the computational function from separate bulky CMOS analog circuits and integrates it directly into the non-volatile memory array structure. The memory cells themselves perform the analog multiplication operations, eliminating the need for external CMOS circuitry and significantly reducing the overall device area while maintaining energy efficiency
Solution Approach 2:
The patent merges the storage function and computation function into a single integrated structure. The non-volatile memory array simultaneously stores weight values and performs analog matrix multiplication operations, combining what were previously separate functions implemented by bulky CMOS circuits into a unified compact structure that achieves both energy efficiency and small footprint
3Use of energy by moving object
If non-volatile memory arrays are used to store weight data in analog form for analog neural network operation, then energy efficiency is improved, but the system cannot operate in digital mode, reducing versatility
Solution Approach 1:
The patent designs the non-volatile memory array system with universal input and output circuitry that can handle both analog and digital operations. The same memory array infrastructure supports analog neural network computations when programmed with analog weight values, while also enabling digital memory operations when programmed with digital data, making the system versatile for multiple operational modes without requiring separate hardware systems
Data Source
Figure 1
Figure 2
Figure 3
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.