Hardware researcher Kate has reportedly used voltage fault injection to bypass factory locks on the NVIDIA CMP 170HX mining card. The technique boosts usable video memory from 8 GB to 80 GB while activating disabled compute units. Modifying a single card remains demanding, requiring between one and 48 hours of hardware work and manual tuning.
The NVIDIA CMP 170HX shares the same GA100 silicon found inside the enterprise A100 accelerator. NVIDIA physically places roughly 80 GB of HBM2e memory on the board, then restricts the hardware for cryptocurrency mining. The company uses eFuse configurations to lock down memory partitions, streaming multiprocessors, PCIe lanes, and NVLink capabilities. Most factory units expose only 8 GB of memory to software, though rare variants ship with 10 GB.
Kate interrupts this protection during startup. By briefly disturbing the power delivery at boot, the voltage fault injection forces the chip to overwrite protected fuse values and restore the disabled silicon. She tested the method across ten cards; nine unlocked the full 80 GB pool, and one capped at 64 GB because of a physically defective memory chip.
The unlocked hardware performs well in compute workloads. In a half‑precision floating‑point benchmark it reached about 85.9 trillion operations per second, versus roughly 72.5 trillion on a reference A100 accelerator in the same test. Still, the process carries real hurdles. Every card needs custom hardware modifications and individual calibration, and the eFuse changes cannot be reversed.

