A researcher says she recovered the untapped memory and computing power of an old Nvidia mining card for artificial intelligence. In a computing test, it even outperforms an A100, a model designed for data centers.
A mining card with 80 GB of hidden memory
Security researcher Kate says she converted an old Nvidia card designed to mine Ethereum into a powerful artificial intelligence tool. In one of her tests, the modified Nvidia CMP 170HX even outperforms the A100, a card from the manufacturer designed for data centers.
Both models are based on the same chip, making this conversion possible. But Nvidia disabled much of the mining card's capabilities, in particular to adapt it for this use.
According to Kate's research, the model sold with 8 GB of memory actually has about 80 GB. This additional memory is physically present, but inaccessible in the original configuration. It could be particularly useful for running artificial intelligence models, which need to load large amounts of data.
Tools already made it possible to recover some of this memory. The researcher went further by also bypassing the hardware restrictions that limited computing power.
A brief voltage drop to bypass the lock
Nvidia differentiates its products in part using electronic fuses built into the chip. Their configuration determines which functions are accessible and which remain disabled.
Kate describes a weakness in this system: certain modifications can reactivate locked capabilities. The cards as sold are normally protected against these modifications.
To bypass this protection, she causes a very brief drop in the voltage supplying the chip. At the right moment, this disruption is enough to evade the protection and allow the necessary modifications.
The researcher uses a small electronic board that she says she designed with the help of an artificial intelligence model. The process nevertheless requires patience: according to her tests, it takes between one and 48 hours of attempts per card.
Faster than an A100 in a computing test
After modification, the mining card can use more compute blocks than a commercially available A100. In the test published by Kate, it therefore performs about 18% more operations per second.
This result does not, however, mean the card will be faster in every artificial intelligence application. It concerns a specific type of computation and does not demonstrate a general advantage for training or running models.
Results also differ from one card to another: of the ten cards examined, nine provided 80 GB of usable memory. The last was limited to 64 GB.
Kate published her research on November 10, after another researcher independently discovered the same vulnerability. She says she notified Nvidia's security team before publication.
