GPU Sort Middle Architecture for Parallel Workload Distribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current graphics processing units (GPUs) face limitations in efficiently processing graphics data due to reliance on fixed function computational units and lack of parallel processing capabilities for general-purpose operations, which hinders performance in tasks like machine learning and pattern analysis.
Innovation Solution
A GPU architecture with a parallel processing unit that includes a scheduler, processing clusters, and a memory crossbar, enabling efficient distribution of workloads across multiple processing clusters and supporting general-purpose parallel compute operations, including graphics processing, machine learning, and pattern analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If fixed function computational units are used in graphics processors, then graphics processing operations can be performed, but the processor cannot support a wider variety of operations for processing vertex and fragment data
Solution Approach 1:
The patent implements a general-purpose graphics processing unit with a unified architecture that can execute both traditional graphics operations and general-purpose compute tasks. The processing clusters contain execution units capable of performing diverse operations including vertex processing, fragment processing, and machine learning workloads, allowing a single hardware design to serve multiple functions rather than requiring separate dedicated units for each operation type
2Productivity
If parallel processing is implemented to increase performance, then processing efficiency improves, but the architecture complexity increases
Solution Approach 1:
The patent divides the processing system into multiple independent processing clusters, each containing execution units, register files, and load/store units. This segmentation allows parallel execution of multiple threads and workloads simultaneously while maintaining modular architecture that simplifies design and management. Each processing cluster can operate independently, enabling scalable parallel processing without requiring complete redesign of the entire system
Solution Approach 2:
The patent implements nested memory hierarchies where processing clusters access shared memory through a memory crossbar structure. The memory system is organized with multiple ports that can simultaneously serve different processing clusters, creating a nested access pattern that enables efficient parallel memory operations without requiring separate memory subsystems for each cluster, thereby reducing overall architectural complexity
3Productivity
If multiple processing clusters are used to maximize parallel processing, then throughput increases, but data distribution and coordination become more complex
Solution Approach 1:
The patent introduces a memory crossbar as an intermediary structure between processing clusters and shared memory. This crossbar provides multiple ports that can simultaneously connect to different processing clusters, enabling efficient data distribution without requiring complex routing logic. The crossbar acts as a mediator that handles data routing and coordination between multiple clusters, simplifying the data distribution problem while maintaining high throughput parallel processing capability
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
Methods, systems and apparatuses may provide for technology that generates, with a first graphics processing unit, geometric attributes associated with a first primitive. The technology determines that the first primitive is associated with a second graphics processing unit based on a first screen space tile associated with the first primitive, and transfers the geometric attributes of the first primitive from the first graphics processing unit to the second graphics processing unit based on the first primitive being associated with the first screen space tile.