NOR FLASH Memory-Based CNN System for Real-Time Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep convolutional neural networks face challenges with large training data and weights, leading to non-real-time data processing and high hardware costs due to their implementation on traditional von Neumann architecture.
Innovation Solution
A memory-based convolutional neural network system utilizing NOR FLASH crossbars for convolutional kernel values and synapse weights, integrating storage and computation to facilitate faster processing and lower hardware costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep convolutional neural networks are implemented on traditional von Neumann architecture, then the network can process data using CPU and GPU, but the processing time increases significantly and hardware costs increase due to frequent data transfer between memory and processor
Solution Approach 1:
The patent merges memory and computing functions into a single integrated device. NOR FLASH memory units directly perform convolution operations without requiring data transfer to separate processors. The crossbar circuit integrates weight storage, multiplication, and accumulation functions within the memory array itself, eliminating the von Neumann bottleneck of separate memory and processing units.
Solution Approach 2:
The patent replaces the mechanical data transfer process between memory and CPU/GPU with direct in-memory computation. Instead of physically moving data through buses and interfaces, the computation occurs where the data is stored, substituting the mechanical transfer system with an in-place computational system using NOR FLASH crossbar circuits.
2Adaptability or versatility
If deep convolutional neural networks store large numbers of weights (up to 10^8) in traditional memory, then the network model can be complex and accurate, but the hardware costs increase significantly
Solution Approach 1:
The patent combines weight storage and computation into the NOR FLASH memory structure. The same physical memory cells that store weights also perform the multiplication operations during inference, eliminating the need for separate high-speed memory and processing units. This integration dramatically reduces hardware costs while maintaining the ability to store large numbers of weights.
Solution Approach 2:
The patent uses the NOR FLASH memory structure itself as the computational substrate rather than copying weights to separate processing units. The memory array's physical structure is utilized to perform computations, replacing the need for duplicate hardware systems for storage and processing.
3Reliability
If traditional memory systems are used to store convolution kernel values and synapse weights, then data can be stored, but real-time processing cannot be achieved due to the von Neumann architecture bottleneck
Solution Approach 1:
The patent merges storage and computation into the NOR FLASH crossbar structure. Convolution kernel values and synapse weights are stored in the same memory units that perform the computations, enabling real-time processing by eliminating data transfer delays between separate memory and processing components.
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
The system achieves real-time data processing and reduces hardware costs by converting software-simulated weights into NOR FLASH threshold characteristics, enabling efficient layer-by-layer calculations and classification.
Implementation Method 1
convolution kernel values or synapse weights are obtained by offline training in a software-simulated environment and are then converted into a threshold characteristic of the NOR FLASH units
Data Source
AI summary
A memory-based CNN, includes an input module, a convolution layer circuit module, a pooling layer circuit module, an activation function module, a fully connected layer circuit module, a softmax function module and an output module, convolution kernel values or synapse weights are stored in the NOR FLASH units; the input module converts an input signal into a voltage signal required by the convolutional neural network; the convolutional layer circuit module convolves the voltage signal corresponding to the input signal with the convolution kernel values, and transmits the result to the activation function module; the activation function module activates the signal; the pooling layer circuit module performs a pooling operation on the activated signal; the fully connected layer circuit module multiplies the pooled signal with the synapse weights to achieve classification; the softmax function module normalizes the classification result into probability values as an output of the entire network.