Evaluating low-level software-based hardening techniques for configurable GPU architectures

The high processing power of GPUs makes them attractive for safety-critical applications, where transient effects are a major concern, and resilience must be enforced without compromising performance. Configurable softcore GPUs are a recent technology that allows detailed reliability assessment capable of bringing directions to the design of reliable GPU applications. This work investigates the reliability of the register files and the pipeline of a softcore GPU under radiation-induced faults. It proposes software-based fault tolerance techniques to mitigate errors. Faults are simulated at the register transfer level in four case-study algorithms, and the Architectural Vulnerability Factor (AVF) and Mean Workload to Failure (MWTF) are checked over different GPU configurations. Results indicate that software-based techniques efficiently reduce AVF. In terms of MWTF, results show that the best cases depend on an optimized balance between GPU configuration, application runtime, and AVF.