Leveraging Algorithm-based Fault Tolerance for Propagation Error Detection in NPUs
Sihyung Kim, Seokin Hong · 2024
As Deep Neural Networks (DNNs) advance, the need for fault-tolerant Neural Processing Units (NPUs) becomes crucial to prevent system failures in safety-critical applications. Transient faults caused by external disturbances like high-energy neutrons or alpha particles can significantly impact NPU reliability. Algorithm-Based Fault Tolerance (ABFT) uses checksums to detect computation errors efficiently at a low cost but has limitations in addressing propagation errors during data propagation. This paper proposes PED-ABFT, a technique that leverages checksums to detect both propagation and computation errors. PED-ABFT enhances error correction by distinguishing between these errors. Our experimental results show PED-ABFT improves error correction coverage by 1.46x with minimal hardware overhead.