Using voltage glitching to unleash the A100 in my mining card

Voltage fault injection on NVIDIA’s GA100 GPU can bypass protections in the fuse controller and program fuses that normally cannot be changed. Applied to an 8 GB CMP 170HX mining card, the technique exposes much of the hardware that the card’s product configuration hides: the card contains roughly 80 GB of HBM, and the unlocked version can use 64-80 GB and enable 120-128 streaming multiprocessors, compared with 108 in a retail A100. The author reports 85.858 TFLOP/s of FP16 throughput - about 1.18 times the measured retail A100 result - because more functional compute units are active.

The investigation explains how GA100’s eFuses control security, product features, and defect management. Regular fuses are difficult to reverse, but “fuseless” fuse records encode option values in a replayable format, so writing new records can replace earlier settings. NVIDIA normally blocks programming these records; a carefully timed voltage glitch disrupts the fuse controller’s protections, enabling the writes. This exposes a design weakness: permanently blowing fuses can make a device less restricted rather than more secure. The method also appears capable of restoring CMP-disabled features such as NVLink, PCIe Gen4, ECC, and multi-instance GPU support; NVDEC, NVJPG, and MIG were not specifically tested. The findings were disclosed to NVIDIA before publication, and the work builds on earlier research into GA100 fuse behavior.