Memcomputing Memory Device for High-Speed Neural Network Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network implementations require significant computational resources, and there is a need for a memory device that can efficiently support high-speed calculations, such as convolution operations, while indicating the signal with the greatest or smallest reference value.
Innovation Solution
A memory device comprising a memory controller, a calculation memory, and a functional circuit, where the calculation memory uses resistors to represent weight values for neural network synapses and outputs signals with reference values, and the functional circuit identifies the signal with the greatest or smallest reference value among the output signals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If neural network calculations are implemented using traditional computing hardware, then calculation accuracy is maintained, but computing speed and efficiency are insufficient for high-speed neural network operations
Solution Approach 1:
The patent replaces traditional electronic computing systems with a memcomputing system that uses magnetic memory cells and spin-transfer torque mechanisms. This substitution enables parallel processing of neural network operations, achieving high-speed computing while maintaining the necessary computational complexity through specialized hardware architecture designed for neural network workloads
Solution Approach 2:
The patent combines memory and computing functions into a unified memcomputing system. By integrating magnetic memory cells with neural network calculation capabilities, the system eliminates the traditional separation between memory and processing units, enabling simultaneous data storage and high-speed neural network computations in a single integrated architecture
2Reliability
If more computational resources are allocated to neural network operations, then calculation accuracy and completeness improve, but energy consumption and resource usage increase significantly
Solution Approach 1:
The patent utilizes changes in magnetic parameters (spin states, magnetization directions) to perform neural network calculations. By leveraging magnetic field strength variations and spin-transfer torque effects, the system achieves accurate computations through physical parameter changes rather than traditional electronic operations, reducing energy consumption while maintaining calculation reliability
Solution Approach 2:
The memcomputing system performs calculations using the inherent magnetic properties of its memory cells. The magnetic moments and spin states naturally interact to compute neural network operations without requiring additional energy-intensive processing steps, enabling the system to serve its own computational needs efficiently while maintaining high accuracy
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Enables efficient implementation of neural network calculations, like convolution, by effectively determining the signal with the greatest or smallest reference value, thereby supporting high-speed neural network operations.
Implementation Method 1
The calculation memory may include a number of memory cells C11-Cmn, wherein m and n are positive integers. Each of the memory cells may include a resistor. The resistance of each of the resistors represents a weight value, which may be used for implementing a Synapse of a neutral network.
Data Source
AI summary
An embodiment of the present invention discloses a memory device. The memory device includes a memory controller, a calculation memory and a functional circuit. The calculation memory is coupled to the memory controller, and is configured to receive a plurality of first signals to output a plurality of second signals. Each of the second signals has a reference value. The functional circuit is coupled to the calculation memory, and is configured to indicate the second signal which has the greatest or the smallest reference value among the second signals.