GPU Partitioning With Matched Functional Throughput
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data processing systems face challenges in efficiently managing multiple independent data processing operations, such as those required for vehicle displays and ADAS, due to complexities in time-sharing GPUs or the high cost of separate units, which can lead to resource inefficiencies and performance bottlenecks.
Innovation Solution
A data processing system with multiple graphics processing units that can operate independently or in combination, utilizing functional units with varying throughputs to match the performance of different sets of graphics processing units, allowing for flexible partitioning and optimal resource allocation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single graphics processing unit is time shared between different data processing functions, then resource utilization is improved, but system complexity and error probability increase while execution independence deteriorates
Solution Approach 1:
The system segments the graphics processing unit into multiple independent partitions, where each partition can be independently allocated to different data processing functions. This segmentation allows multiple functions to execute simultaneously in separate partitions rather than time-sharing a single unified GPU, thereby maintaining resource utilization while reducing complexity and enabling execution independence.
Solution Approach 2:
The patent introduces a spatial dimension to resource allocation by dividing the GPU into multiple partitions that can operate in parallel. Instead of temporal multiplexing (time-sharing), the system uses spatial multiplexing (partitioning), adding a dimensional aspect to resource management that resolves the contradiction between utilization and independence.
2Reliability
If completely separate graphics processing units are provided for each data processing function, then execution independence is improved, but hardware cost and component quantity increase
Solution Approach 1:
The system merges multiple graphics processing units into a single unified GPU that contains multiple independent partitions. These partitions can be independently allocated to different data processing functions, providing execution independence similar to separate GPUs while reducing the total component count and hardware cost.
Solution Approach 2:
The unified GPU is designed with multi-functionality, where a single device can serve multiple data processing functions through its partitioned architecture. Each partition can be independently configured and allocated to different functions, allowing one GPU to replace multiple dedicated GPUs while maintaining functional independence.
3Ease of manufacture
If fixed resource division is implemented at SoC creation time, then resource allocation simplicity is improved, but adaptability to different performance requirements deteriorates
Solution Approach 1:
The system implements dynamic resource allocation where the partitioning and allocation of GPU resources can be changed at runtime based on performance requirements. Unlike fixed resource division at manufacturing time, this dynamic approach allows the system to adapt to different workload demands while maintaining simple underlying hardware architecture.
Solution Approach 2:
The patent enables parameter changes in resource allocation by allowing the number, size, and configuration of GPU partitions to be dynamically adjusted. This flexibility permits the system to adapt to varying performance requirements without changing the physical hardware structure, resolving the contradiction between manufacturing simplicity and operational adaptability.
Data Source
AI summary
A data processing system that comprises plural processing units is disclosed. The system includes functional units, the functional units having different processing capacities. A set of one or more processing units can operate in combination with one of the functional units according to a processing capacity required for the set of one or more processing units.