GPU Thread Error Isolation Through Exclusive Resource Assignment

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Processor resource assignment methods, particularly in GPUs, lead to resource idle time and difficulty in differentiating faults between multiple parallel operating applications, resulting in inefficient operation and inaccurate fault reporting.

Innovation Solution

Implementing a system that allows for exclusive assignment of processing resources to applications, enabling selective stopping of threads based on errors, and maintaining a tracking system to isolate fault reporting, thereby ensuring accurate fault identification and minimizing resource downtime.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If multiple applications share processing resources in parallel, then resource utilization increases, but fault differentiation between applications becomes difficult

Engineering Contradiction:
Improveresource utilizationVSAvoidfault differentiation accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent segments processing resources into application-specific resource sets, where each application is assigned exclusive access to its designated resources. This segmentation allows faults to be isolated to specific applications, enabling accurate fault differentiation while maintaining high resource utilization through dynamic reassignment when applications complete or fail.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If processing resources are exclusively assigned to applications, then fault reporting accuracy improves, but resource idle time increases

Engineering Contradiction:
Improvefault reporting accuracyVSAvoidresource idle time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent implements dynamic resource assignment where applications are exclusively assigned to processing resources during their execution, ensuring accurate fault reporting. When an application completes or encounters a fault, the system dynamically reassigns those resources to other waiting applications, minimizing idle time while maintaining exclusivity during active execution.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system maintains continuous useful action by implementing a queue of waiting applications that can immediately utilize freed resources. This ensures that processing resources transition smoothly from one application to the next without significant idle periods, while each application enjoys exclusive access during its execution for accurate fault reporting.

Inventive Principle:
Principle #20Continuity of useful action

3Productivity

If processing resources are shared between multiple applications, then resource efficiency improves, but resource downtime increases when faults occur

Engineering Contradiction:
Improveresource efficiencyVSAvoidresource downtime
Core Design Contradiction:
ProductivityVSDuration of action of moving object

Solution Approach 1:

By segmenting resources into application-specific sets with exclusive assignment, the patent enables precise fault isolation. When a fault occurs in one application, only that application's resource set is affected and can be independently managed, allowing other applications to continue using their designated resources without interruption, thus minimizing overall resource downtime.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system ensures continuity of useful action by maintaining queues of ready applications that can immediately take over freed resources. When faults occur, the exclusive assignment model allows for rapid reassignment of affected resources to waiting applications, reducing the duration of resource downtime while maintaining high overall efficiency.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP4650961A1Identifying thread errors
Publication Date: 2025.11.19 NVIDIA CORP
  • EP4650961A1 patent drawingFigure 1
  • EP4650961A1 patent drawingFigure 2
  • EP4650961A1 patent drawingFigure 3

AI summary

Apparatuses, systems, and techniques to perform exclusive assignment of processing resources, during operation, to operating applications to allow for exclusive fault reporting between applications. In at least one embodiment, processors comprising one or more circuits to cause performance of one or more threads corresponding to one or more respective kernels to be selectively stopped based, at least in part, on at least one of the one or more threads encountering an error.