Graphics Processing Units (GPUs) are widely de-ployed and utilized across various computing domains including cloud and high-performance computing. Considering its extensive usage and increasing popularity, ensuring GPU reliability is cru-cial. Software-based reliability evaluation methodologies, though fast, often neglect the complex hardware details of modern GPU designs. This oversight could lead to misleading measurements and misguided decisions regarding protection strategies. This paper breaks new ground by conducting an in-depth examination of well-established vulnerability assessment methods for modern GPU architectures, from the microarchitecture all the way to the software layers. It highlights divergences between popular software-based vulnerability evaluation methods and the ground truth cross-layer evaluation, which persist even under strong protections like triple modular redundancy. Accurate evaluation requires considering fault distribution from hardware to software. Our comprehensive measurements offer valuable insights into the accurate assessment of GPU reliability.

Due to extensive deployment and heavy usage of GPUs, ensuring the reliability of such devices is crucial. Current software-based reliability evaluation methodologies, albeit fast, often neglect the intricate hardware complexities of modern GPU designs. This oversight could result in misleading measurements and misguided decisions regarding protection strategies. This work breaks new ground by examining well-established vulnerability assessment methods for modern GPU architectures, from the microarchitecture all the way to the software layers. It highlights divergences between popular software-based vulnerability evaluation methods and the ground truth cross-layer evaluation (which, as we show, holds even when strong protection like triple modular redundancy is employed); accurate evaluation requires considering fault distribution from hardware to software. Our comprehensive measurements offer valuable insights into accurately assessing GPU reliability.

Pre-silicon reliability evaluation of processors is usually performed at the microarchitecture or at the software level. Recent studies on CPUs have, however, shown that software level approaches can mislead the soft error vulnerability assessment process and drive designers towards wrong error protection decisions. To avoid such pitfalls in the GPUs domain, the availability of microarchitecture level reliability assessment tools is of paramount importance. Although there are several publicly available frameworks for the reliability assessment of GPUs, they only operate at the software level, and do not consider the microarchitecture. This paper aims at accurate microarchitecture level GPU soft error vulnerability assessment. We introduce gpuFI-4: a detailed microarchitecture-level fault injection framework to assess the cross-layer vulnerability of hardware structures and entire GPU chips for single and multiple bit faults, built on top of the state-of-the-art simulator GPGPU-Sim 4.0. We employ gpuFI-4 for fault injection of soft errors on CUDA-enabled Nvidia GPU architectures. The target hardware structures that our framework analyzes are the register file, the shared memory, the LI data and texture caches and the L2 cache, altogether accounting for tens of MBs of on-chip GPU storage. We showcase the features of the tool reporting the vulnerability of three Nvidia GPU chip models: two different modem GPU architectures – RTX 2060 (Turing) and Quadro GV100 (Volta) – and an older generation – GTX Titan (Kepler), for both single-bit and triple-bit fault injections and for twelve different CUDA benchmarks that are simulated on the actual physical instruction set (SASS). Our experiments report the Architectural Vulnerability Factor (AVF) of the GPU chips (which can be only measured at the microarchitecture level) as well as their predicted Failures in Time (FIT) rate when technology information is incorporated in the assessment.