Crossbar Circuit Switching Between Regular and Depthwise Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing crossbar circuits are limited in their ability to efficiently perform both regular and depth-wise convolution operations, which are crucial for advanced neural network computations.
Innovation Solution
A crossbar-based apparatus with multiple crossbar arrays and select circuits that can switch between regular and depth-wise convolution operations by selectively programming cross-point devices using multiplexers and control signals, allowing for efficient performance of both types of convolutions using the same crossbar architecture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single crossbar array is used for both regular and depth-wise convolution operations, then device complexity is reduced, but the circuit cannot efficiently perform both operation types simultaneously
Solution Approach 1:
The patent implements multi-functionality by configuring a single crossbar array to perform both regular and depth-wise convolution operations through selective device activation. The crossbar array is designed with word lines, bit lines, and select lines that can be dynamically controlled to enable different operation modes, allowing one hardware structure to serve multiple computational purposes without requiring separate dedicated circuits for each convolution type
Solution Approach 2:
The patent applies dynamics by introducing selectable configurations through control signals that activate different subsets of cross-point devices during operation. The select lines and associated circuitry enable dynamic reconfiguration of the crossbar array's functional state, allowing the system to switch between regular and depth-wise convolution modes by selectively enabling or disabling specific device groups based on the required operation type
2Productivity
If separate crossbar arrays are used for regular and depth-wise convolution operations, then operational efficiency is improved, but device complexity and area increase
Solution Approach 1:
The patent eliminates the need for separate dedicated crossbar arrays by designing a universal crossbar structure that can be dynamically configured for different convolution types. This universal design reduces the total hardware area required while maintaining the ability to efficiently perform both regular and depth-wise operations through temporal multiplexing and selective device activation rather than spatial duplication
3Adaptability or versatility
If the crossbar array is reconfigured for different convolution operations, then adaptability is improved, but switching time and operational speed are reduced
Solution Approach 1:
The patent minimizes reconfiguration overhead by implementing a dynamic selection mechanism that activates pre-positioned subsets of cross-point devices rather than physically reconfiguring the entire array. The select lines and control circuits enable rapid mode switching by simply changing which devices are conductive, allowing the system to transition between operation types with minimal delay and without disrupting the overall array structure
Solution Approach 2:
The patent applies preliminary action by pre-organizing the cross-point devices into groups that can be selectively activated for different operation types. The crossbar array is designed with word lines, bit lines, and select lines that are预先 configured to support both regular and depth-wise operations, allowing the system to switch modes by simply enabling the appropriate pre-prepared device subsets rather than performing complex reconfiguration during operation
Data Source
AI summary
In some embodiments, an apparatus for performing convolution operations is provided. The apparatus may include multiple crossbar arrays and select circuits. The select circuits are configured to select a first plurality of cross-point devices and a second plurality of cross-point devices in response to receiving a control signal indicating that a regular convolution is to be performed, and to select the first plurality of cross-point devices and a third plurality of cross-point devices in response to receiving a control signal indicating that a depthwise convolution is to be performed. The first plurality of cross-point devices is connected to a first plurality of word lines and a first bit line. The second plurality of cross-point devices is connected to the first plurality of word lines and a second bit line. The third plurality of cross-point devices is connected to a second plurality of word lines and the second bit line.