Researcher Unlocks CMP 170HX and Obtains A100-Level Card

In a published study, Kate described a method to bypass the hardware limitations of the NVIDIA CMP 170HX mining GPU. She used controlled voltage glitches to gain access to the fuse controller on the GA100 die.

The CMP 170HX, with 8 GB of available memory, is actually equipped with a larger volume of HBM, which the manufacturer disables at the chip level. After modification, Kate was able to enable up to 80 GB of memory. According to her data, this result turns the card into an accelerator functionally close to the A100 for artificial intelligence tasks.

For the attack, the researcher connected an additional microcontroller-based circuit to the board. She intercepted voltage control signals, after which repeated reading of the fuse states allowed the controller to be switched into a writable mode. This enabled Kate to unlock compute units, memory, PCIe Gen4, NVLink, ECC, and Multi-Instance GPU mode.

The researcher separately noted that physically defective compute units cannot be enabled without verification, as irreversible fuses are used for them. She did not test video codec functions and some other capabilities.

In tests, the modified card demonstrated approximately 20% higher FP16 FMA performance than the standard A100. Kate attributes this to the fact that more compute units are active on the unlocked chips, whereas in production A100s, NVIDIA left 108 such units to improve chip yield.

Prior to publication, Kate reported the vulnerability to the NVIDIA team, and cited the independent discovery of the same method by another researcher as the reason for disclosure.