Context-Dependent Deep Learning with Clustered Thread Scheduling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning algorithms face challenges in efficiently utilizing parallel processing capabilities of general-purpose graphics processing units (GPGPUs) for tasks such as deep neural networks, particularly in managing thread synchronization and resource allocation.
Innovation Solution
Implementing a parallel processor architecture with a scheduler that allocates work to processing clusters using various scheduling algorithms, enabling efficient parallel processing of machine learning operations, including graphics processing and general-purpose compute operations, while utilizing a unified memory design for seamless data access.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If parallel processing is used to improve machine learning computation speed, then processing throughput increases, but thread synchronization complexity increases
Solution Approach 1:
The patent divides the parallel processing system into multiple independent processing clusters, each capable of autonomous operation. This segmentation allows threads to be organized into discrete groups that can be managed independently, reducing the overall synchronization complexity while maintaining high parallel throughput for machine learning computations.
Solution Approach 2:
The patent introduces a scheduler as an intermediary component that manages thread allocation and coordination between processing clusters. This mediator handles the complex synchronization requirements by centralizing thread management, allowing individual clusters to operate with reduced synchronization overhead while maintaining system-wide coordination.
2Measurement precision
If more data is processed to improve model accuracy, then training quality increases, but processing time increases
Solution Approach 1:
The patent implements data preprocessing and feature extraction pipelines that prepare training data in advance before it reaches the main processing stage. By performing preliminary actions on the data, the system can process larger datasets more efficiently, improving model accuracy without proportionally increasing training time.
Solution Approach 2:
The patent employs continuous data streaming and iterative processing approaches where data is fed continuously through the network in optimized batches. This allows the system to process large datasets without idle time between processing stages, maintaining continuous useful action that reduces overall processing time while preserving 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 receive a plurality of data inputs for training a neural network, wherein the data inputs comprise training data and weights inputs; represent the data inputs in a first form; and represent the weight inputs in a second form. Other embodiments are also disclosed and claimed.