GPU Diverse Redundancy for Fault Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current GPU subsystems face challenges in ensuring data content integrity for safety-critical applications due to transient or permanent failures, lack of built-in runtime diagnostics, and the expense of error correcting code (ECC) memory, making it difficult to detect operational faults without additional hardware components.
Innovation Solution
The GPU subsystem performs the same graphics operations twice in different rendering modes, comparing the results to detect operational faults and issue warnings, enabling concurrent and online testing without entering a dedicated test mode, and utilizing different hardware components and buses to detect faults that may not be apparent in a single mode.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the GPU subsystem uses ECC memory to ensure data integrity, then reliability is improved, but device complexity and cost increase
Solution Approach 1:
The patent implements redundancy by performing the same graphics operation twice using different rendering modes (e.g., rasterization mode and ray tracing mode) to produce duplicate results that are then compared for consistency. This copying approach detects faults without requiring ECC memory, resolving the contradiction between reliability and device complexity.
Solution Approach 2:
The GPU subsystem uses its existing multi-functional rendering capabilities to perform both normal graphics operations and fault detection operations. By leveraging the universality of the rendering pipeline to execute multiple rendering modes, the system achieves reliability improvement without adding dedicated fault detection hardware.
2Measurement precision
If the GPU subsystem enters a dedicated test mode to detect faults, then measurement precision is improved, but productivity decreases due to system downtime
Solution Approach 1:
The patent enables continuous fault detection by performing redundant rendering operations alongside normal graphics processing. The system continuously compares results from different rendering modes without interrupting productivity, as the fault detection is integrated into the normal operation flow rather than requiring separate test modes.
Solution Approach 2:
The GPU subsystem performs self-diagnosis by comparing its own rendering outputs from different modes. This self-service approach allows the system to detect faults autonomously during normal operation without requiring external testing equipment or dedicated test modes, maintaining both measurement precision and productivity.
3Ease of operation
If the GPU subsystem uses identical rendering modes for redundancy, then ease of operation is improved, but reliability decreases as the same faults go undetected
Solution Approach 1:
The patent employs asymmetric rendering modes (e.g., combining rasterization with ray tracing, or using different anti-aliasing methods) for the redundant operations. This asymmetry ensures that different types of faults manifest differently in each mode, enabling comprehensive fault detection while maintaining operational simplicity through automated mode selection and result comparison.
Data Source
Figure 1
Figure 2
Figure 3A~3B
AI summary
A graphics processing unit (GPU) of a GPU subsystem of a computing device operates in a first rendering mode to process graphics data to produce a first image. The GPU operates in a second rendering mode to process the graphics data to produce a second image. The computing device detects whether a fault has occurred in the GPU subsystem based at least in part on comparing the first image with the second image.