Deep Learning Network Scheduling via Parallel GPU Execution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning processing technologies face inefficiencies in training and deploying deep neural networks due to limitations in parallel processing capabilities, particularly in handling large datasets and complex computations.
Innovation Solution
The integration of a general-purpose graphics processing unit (GPU) with a parallel processor architecture, utilizing SIMT (Single Instruction, Multiple Thread) and high-speed interconnects like PCIe or NVLink, to accelerate machine learning operations and parallelize computations across multiple GPUs, enabling efficient training and deployment of deep neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If parallel processing is used to accelerate machine learning computations, then processing speed and efficiency are improved, but device complexity increases
Solution Approach 1:
The patent segments the deep learning network into multiple layers and operations that can be executed in parallel across multiple GPU cores. Each layer or operation is divided into independent tasks that can be processed simultaneously, enabling parallel execution while maintaining manageable complexity through structured decomposition.
Solution Approach 2:
The patent implements a universal scheduling framework that can handle different types of deep learning operations (convolution, fully connected, pooling, etc.) through a common parallel processing architecture. This multi-functional approach allows the same hardware infrastructure to execute diverse machine learning workloads without requiring specialized dedicated hardware for each operation type.
2Measurement precision
If larger datasets and more complex computations are processed, then model accuracy and capability are improved, but training time and computational resources increase
Solution Approach 1:
The patent performs preliminary actions by pre-processing and organizing training data into formats optimized for parallel processing before feeding it to the GPU. Additionally, the scheduling framework prepares and configures parallel execution pipelines in advance, so that when training begins, computations can immediately proceed in parallel without setup delays, reducing overall training time while maintaining accuracy.
Solution Approach 2:
The patent ensures continuous utilization of GPU resources by implementing overlapping computation and data transfer operations. While one set of computations is being executed, data for subsequent computations is being prepared or transferred, eliminating idle time and maintaining continuous productive action throughout the training process, thereby reducing total training time without compromising model accuracy.
Data Source
AI summary
In an example, an apparatus comprises a plurality of execution units comprising and logic, at least partially including hardware logic, to traverse a solution space, score a plurality of solutions to a scheduling deep learning network execution, and select a preferred solution from the plurality of solutions to implement the deep learning network. Other embodiments are also disclosed and claimed.