GPU Canary: Border Inputs as a Two-Second Integrity Check for Silent Data Corruption in Deployed Language Models

A GPU with a silent hardware fault returns wrong arithmetic without raising an error. Production studies report that such faults are routine at fleet scale, that three quarters of affected cards pass burn-in, and that vendor diagnostics recover about 70% of them at several hours per card. The cheapest online check, redundant execution with bitwise comparison, is blind by construction to the common case of a unit that is wrong the same way on every run. We show that a language model already loaded on a GPU can act as its own integrity probe. We mine 64 border inputs, prompts on which the model's top two next-token logits are nearly tied, so that any perturbation of the arithmetic flips the greedy decision, record their answers while the hardware is healthy, and later replay them and count mismatches. On an NVIDIA L4 in fp16 the reference is bit-stable across reruns, and a scan of 64 probes takes 2.5 s. Against software-injected faults reproducing the patterns reported in production (a value zeroed, a sign flipped, a bit flipped, a relative bias of 1e-3) at one corruption per hundred layer forwards, the probes flag 10 of 10 scans for every pattern, while 32 ordinary prompts flag between 0 and 8 of 10 and never see the 1e-3 bias. Against persistent faults fixed at a random site, the probes flag 8 of 8 sites for every pattern on two models, redundant execution flags 0 of 8, and ordinary prompts 0 or 1 of 8. Under the RTL-derived permanent-fault model of LLM-PRISM, hard tensor-core defects are flagged at 15 of 16 sites by the probes, 1 of 16 by ordinary prompts, and 0 of 16 by redundant execution, while intermittent defects favour redundant execution; the two checks are complementary. The method is the border-input idea of recent model-change-detection work, redirected at the hardware. No genuinely defective GPU has been tested; all faults are injected, and we state the limits plainly. Code, probe bank, reference signatures and run logs: https://github.com/rishvaiyer/gpu-canary