Convolution Arithmetic Circuit for CNN SoC Efficiency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Convolutional Neural Networks (CNNs) face significant computational challenges when implemented on resource-limited built-in SoC devices due to the large number of convolution arithmetic operations required, making it difficult to perform image recognition tasks efficiently.

Innovation Solution

The proposed solution involves generating an integration coefficient table that reduces the number of convolution arithmetic operations by performing cumulative additions across channels in parallel, using a convolution arithmetic processing circuit and an integration coefficient table generation device to integrate coefficients from multiple channels, allowing for efficient CNN operation on limited resource devices.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If CNN performs convolution arithmetic operations for each channel separately, then image recognition accuracy is maintained, but the operation amount becomes extremely large

Engineering Contradiction:
Improveimage recognition accuracyVSAvoidoperation amount
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent merges convolution operations across multiple channels by generating an integration coefficient table that combines coefficients from multiple channels. This allows the system to perform a single integrated convolution operation instead of separate operations for each channel, significantly reducing the total operation amount while maintaining recognition accuracy through the integrated coefficient information.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent performs preliminary action by pre-calculating and storing integrated coefficients in an integration coefficient table before the actual convolution operation. This pre-processing step consolidates channel-specific coefficients into integrated coefficients that can be used directly during inference, eliminating the need for repeated separate channel operations.

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If CNN is implemented on built-in type SoC devices, then device integration is improved, but resource limitations make it difficult to handle large operation amounts

Engineering Contradiction:
Improvedevice integrationVSAvoidresource constraints
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent merges multiple channel operations into a single integrated operation using the integration coefficient table. This reduction in operation count directly addresses the resource constraints of built-in type SoC devices, making CNN implementation feasible on resource-limited hardware while maintaining the adaptability to run on various device types.

Inventive Principle:
Principle #5Merging (Combining)

3Measurement precision

If multiple convolution operations are performed for each channel, then comprehensive feature extraction is achieved, but processing time increases

Engineering Contradiction:
Improvefeature extraction accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary action by pre-computing integrated coefficients that encapsulate feature extraction information from multiple channels. During actual processing, this pre-computed table allows for rapid convolution operations without sacrificing the comprehensive feature extraction capability, thereby significantly reducing processing time.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent merges the feature extraction operations across channels by using integrated coefficients that combine information from multiple channels. This merging allows the system to achieve comprehensive feature extraction in a single operation rather than requiring separate operations for each channel, thus reducing processing time.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3447682B1Semiconductor device and image recognition system
Publication Date: 2023.05.24 RENESAS ELECTRONICS CORP
  • EP3447682B1 patent drawingFigure 1~2A
  • EP3447682B1 patent drawingFigure 2B~2C
  • EP3447682B1 patent drawingFigure 3

AI summary

A semiconductor device includes an image recognition device having a convolution arithmetic processing circuit. The convolution arithmetic processing circuit includes a coefficient register where coefficients of an integration coefficient table are set, a product calculation circuit that calculates products of an input image and the coefficients, a channel register where a channel number of the integration coefficient table is set, a channel selection circuit that selects an output destination of a cumulative addition arithmetic operation on the basis of the channel number, and a plurality of output registers that store a result of the cumulative addition arithmetic operation. The integration coefficient table is a table where a plurality of input coefficient tables are integrated and the integration coefficient table has a size of N x N. The product calculation circuit can calculate data of N x N all at once.