Researcher uses weight patterns to trace which GPUs trained AI models

Started by Highland Canopy, Yesterday at 09:05 AM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Topic: Researcher uses weight patterns to trace which GPUs trained AI models   Views(Read 73 times)
Active members in this topic:
Highland Canopy(1)

Highland Canopy

A former OpenAI engineer says he can tell which chips were used to train an AI model. The Register reports that Will DePue, now at Unemployment Labs, claims several Chinese labs trained frontier models on Nvidia Blackwell GPUs

He names DeepSeek, Moonshot, MiniMax and Ant Group, with models including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3. His method relies on the way weight initialisation patterns from Philox and cuRAND differ by GPU type and random seed, which he says survive training

He searched more than six trillion combinations of chip layout and seed. He says six models show conclusive evidence, with four matching GB200 or GB300 parts

The Register notes that this shows Blackwell use, not where the chips were located, since cloud access is possible. The findings have not been independently verified, and neither the labs nor Nvidia responded

Whatever the wider context, the forensic technique itself is fascinating. Being able to fingerprint the hardware behind a model from its weights could have many uses, from research to auditing. Independent researchers will want to test whether the method holds up

Could hardware fingerprinting become a standard tool for auditing AI models? And how reliable do you think this kind of analysis is?