PCM Neural Network Device Shared Backward Spike Generators
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural network devices have large circuit areas and high energy consumption due to unnecessary components in input and output amplifiers, which are not optimized for specific functions.
Innovation Solution
A phase change material (PCM)-based neural network device is designed with shared Backward Spike Generators (BSGs) and different components for each layer, reducing unnecessary components and optimizing circuit area and energy usage by synchronizing pulse outputs through control circuits.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional neural network devices use identical structures for input and output amplifiers including Spike Generators, then the device structure is standardized and easier to manufacture, but the circuit area increases and energy consumption rises due to unnecessary components
Solution Approach 1:
The patent applies local quality by configuring input layer neurons with forward pulse drivers and output layer neurons with backward pulse drivers, tailoring each layer's components to its specific functional requirements rather than using a uniform structure throughout, thereby eliminating unnecessary components in each layer
Solution Approach 2:
The patent merges the Spike Generator functionality into the control circuit that is already present in each layer, eliminating the need for separate SG components while maintaining the spike generation function through integrated control, thus reducing overall circuit area
2Device complexity
If conventional neural network devices include Spike Generators in each amplifier, then the device follows a consistent design pattern, but energy consumption increases due to unnecessary components
Solution Approach 1:
The patent extracts the Spike Generator functionality from the amplifier structure and implements it through the control circuit's timing control, removing unnecessary components from each neuron while maintaining the essential spike generation capability through centralized control mechanisms
Solution Approach 2:
The control circuit is designed to perform multiple functions: it controls pulse generation timing, manages forward and backward pulse drivers differently based on layer type, and coordinates neuron activation, thereby replacing the need for dedicated Spike Generators in each amplifier while maintaining design consistency
3Ease of manufacture
If conventional neural network devices use the same amplifier structure for all layers, then manufacturing is simplified, but circuit area and energy consumption increase due to unnecessary components in each layer
Solution Approach 1:
The patent implements local quality by configuring input layer neurons with forward pulse drivers and output layer neurons with backward pulse drivers, optimizing each layer's component structure to match its specific functional requirements and eliminating energy-wasting unnecessary 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 PCM-based neural network device minimizes circuit area and energy consumption by sharing BSGs and using only necessary components for each layer, while synchronizing pulse outputs to enhance efficiency.
Implementation Method 1
a plurality of PCMs connecting between input lines of the input layers and output lines of the output layers
Data Source
AI summary
A phase change material (PCM)-based neural network device according to an embodiment comprises: a plurality of neurons disposed for input layers and output layers, respectively; a plurality of PCMs connecting input lines of the input layers and output lines of the output layers; and at least one backward spike generator (BSG) shared by the plurality of neurons, and generating spike on the basis of an output pulse outputted from each of the neurons of the output layers.