Who cares?
I’m more bothered that they are fingerprining GPUs in such a way to track this.
In they article they explain that it’s technical details of how the hardware functions that leave a fingerprint. Not an intentional tracking method.
Good.
Use the hardware to keep R&D pace with US models while pushing domestic hardware adoption for inference. Gives us free models to run while doing it.
Meh. I just want them to be cheap so that the future is democratized.
I thought this would be nonsense at first, but the methodology is interesting. It seems plausible. And they tested quite a good spread of open weights, like MiMo and LongCat (which usually fly under the radar), all the way back to Yi 34B.
They missed Russia’s LLM, but that’s okay; it’s very obscure.
It’s also “lucky” they found this correlation, and so many recent weights happened to be initialized with PyTorch, on the GPU: very specific conditions.
…Also, I think the FP4 precision is a dead giveaway too, as Hopper wouldn’t be able to train that effectively, right? I assumed they were using at least some smuggled B200s.
And the export restrictions were stupid anyway.
correct me if i am wrong, but isn’t their methodology only able to recover the number of compute units the chip has?
they even admitted themselves:
but we confirmed on a physical MI300X that, limited to 152 or 148 of its units, it reproduces the layouts we associate with Blackwell, so the layout alone does not settle the question.
they did then add they may be able to differentiate by numerical difference in calculation, but that approach was inconclusive for many of the models.
could also be domestic chinese GPUs the researchers had no access to.
