divagr18/gpu-instance-fingerprint
Fingerprint the exact GPU you're running on, even among identical GPUs, from the order its atomic adds land in. 99.1% accuracy, survives reboots, can't be faked, and a default torch.index_add_ call leaks it.
GitHub repository with 27 stars and 5 forks.
Language: Python
Topics: cuda, cupy, device-fingerprinting, fingerprinting, floating-point, gpu, hardware-security, nondeterminism, nvidia, privacy