SHOUT — Silent Data Corruption Hunting and Observation Using Transformers

Silent Data Corruptions (SDCs) have become a major reliability concern in large-scale computing systems, with recent fleet studies reporting up to 5000 Defective Parts per Million (DPM) test escapes and hundreds of thousands of silent corruptions per billion chip-hours. Although these failures are deterministic under particular voltage, thermal, and operand conditions, they remain difficult to capture because traditional defect models do not exercise the exact conditions required to reveal them. To address this challenge, we propose SHOUT methodology in which faults are injected directly into instruction results to hunt for SDC vulnerable patterns. This process produces a large labeled dataset that captures the behavior of real program operations under perturbations and reveals vulnerability trends consistent with those expected in production. Building on this dataset, a transformer based model is fine tuned to identify SDC prone lines directly from source code.1 The model achieves 94% accuracy and an F1-score of 0.973, demonstrating that the proposed method can identify silent error susceptible lines without requiring program execution or hardware access.1Upon acceptance, the SHOUT framework including the dataset, sample programs, instrumentation library, and fault injection results will be made available on request.