Arithmetic-intensity-guided fault tolerance for neural network inference on GPUs

• Main Authors: Kosaian, Jack, Rashmi, K. V.
• Format: Conference Proceeding
• Language: English
• Subjects: arithmetic intensity; Artificial neural networks; Computer systems organization > Dependable and fault-tolerant systems and networks > Redundancy; Fault tolerance; Fault tolerant systems; General and reference > Cross-computing tools and techniques > Reliability; Graphics processing units; High performance computing; Instruction sets; neural networks; Space vehicles
• ISSN: 2167-4337
• DOI: 10.1145/3458817.3476184

Neural networks (NNs) are increasingly employed in safety-critical domains and in environments prone to unreliability (e.g., soft errors), such as on spacecraft. Therefore, it is critical to impart fault tolerance to NN inference. Algorithm-based fault tolerance (ABFT) is emerging as an efficient approach for fault tolerance in NNs. We propose an adaptive approach to ABFT for NN inference that exploits untapped opportunities in emerging deployment scenarios. GPUs have high compute-to-memory-bandwidth ratios, while NN layers have a wide range of arithmetic intensities. This leaves some layers compute bound and others memory-bandwidth bound, but current approaches to ABFT do not consider these differences. We first investigate ABFT schemes best suited for each of these scenarios. We then propose intensity-guided ABFT, an adaptive, arithmetic-intensity-guided approach that selects the most efficient ABFT scheme for each NN layer. Intensity-guided ABFT reduces execution-time overhead by 1.09--5.3× across many NNs compared to traditional approaches to ABFT.