In an Oct. 30 post, divagr1925 introduced a GPU-fingerprinting project claiming that a default PyTorch torch.index_add_ operation can help identify individual GPUs, even among chips of the same model. The project’s test used a deliberately configured input, not just any call to the operation.
The project’s README says its PyTorch test put four contributions into each output bin, using values that ranged from 2^-30 to 2^30 with random signs. It used a fixed seed to give every chip identical input. The README attributes the identifying differences to the order in which atomic additions reach each bin: changing the addition order can change a floating-point result. A denser test, by contrast, classified chips at chance.
On six NVIDIA A40 GPUs, the README reports 9 of 9 correct classifications among chips using the same driver and 78% accuracy across mixed drivers. The headline 99.1% result came from a different experiment: a purpose-built CUDA kernel correctly classified 111 of 112 measurements in a leave-one-out test of 19 identical A40s. It is not the reported accuracy of torch.index_add_.
The README also reports that all 12 measurements taken after a container restart matched the right chip in a four-A40 test. That is a narrower finding than a guarantee that a fingerprint will survive every reboot. In other tests on three A40s, the project reports bit-identical results across chips for a vanilla transformer’s forward logits and embedding gradient, as well as tested matrix multiplication. Those results do not establish that ordinary model inference exposes a GPU’s identity.
PyTorch’s reproducibility documentation provides another limit: identical random seeds do not guarantee reproducibility across releases or platforms. Its deterministic-algorithms setting selects deterministic implementations where available and otherwise raises an error. The documentation’s example shows such an error for a nondeterministic CUDA index_add_ implementation, so enabling the setting is not a blanket promise that the operation will keep running unchanged.
Comments