CNN Training Task Splitting Across Heterogeneous Processors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The inflexibility and speed limitations of existing hardware platforms for training convolutional neural network (CNN) models, particularly due to the inflexible migration and computing speed issues when tasks are migrated or co-computed across different computing devices or processors.
Innovation Solution
A method and apparatus that split the multiply-accumulate operations in CNN model training tasks into multiple multiply-add operations, identify corresponding computing devices for each task, and perform computations on these tasks separately to enhance flexibility and speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If CNN model training tasks are migrated on different computing devices or co-computed by different processors, then computing flexibility is improved, but computing speed is seriously affected due to inflexible task migration and dedicated computing execution granularity
Solution Approach 1:
The patent segments the CNN training task into multiple sub-tasks with different granularities (first granularity for CPU, second granularity for GPU, third granularity for FPGA). This segmentation allows each computing device to process tasks at its optimal granularity level, improving both migration flexibility and computing speed by matching task size to device capabilities.
Solution Approach 2:
The patent implements dynamic task allocation where the computing system can flexibly assign different granularities of sub-tasks to different computing devices based on their capabilities and current workload. This dynamic approach enables efficient load balancing and optimal utilization of heterogeneous computing resources while maintaining high computing speed.
2Productivity
If dedicated and customized computing execution granularity is used for CNN training tasks on different computing devices, then computing efficiency on each device is improved, but task migration and cooperative computing become inflexible
Solution Approach 1:
The patent creates a universal task allocation framework that can handle multiple computing devices (CPU, GPU, FPGA, etc.) with different architectures. The system universally applies the methodology of dividing tasks into multiple granularities across all device types, enabling flexible migration and cooperative computing while maintaining optimal efficiency on each specific device type.
3Measurement precision
If CNN model size is increased to improve accuracy in target detection and recognition, then model accuracy is improved, but hardware platform requirements become increasingly demanding and reach Moore's law bottleneck
Solution Approach 1:
The patent segments large-scale CNN training tasks into smaller sub-tasks that can be distributed across multiple computing devices. This segmentation allows the system to handle large models without requiring a single monolithic hardware platform, thereby avoiding Moore's law limitations while maintaining the ability to train accurate large-scale models through distributed computing.
Data Source
AI summary
A computing method and apparatus for a convolutional neural network model. The method comprises: acquiring a computing model of a training task of a convolutional neural network model (S101); then splitting multiply-accumulate operation in a computing model of a training task of the convolutional neural network model into a plurality of multiply-add operation tasks (S102); confirming a computing device corresponding to each multiply-add operation task according to the correlation between a preset computing model and the computing device (S103); and finally, respectively computing each multiply-add operation task by utilizing the computing device corresponding to each multiply-add operation task (S104). The purposes of improving the flexibility of migration of a CNN model training task on different computing devices or cooperative computing of different processors and improving the computing speed are achieved.