A researcher under the nickname Kate was able to unlock the hidden hardware capabilities of the Nvidia CMP-170HX (8 GB) mining video card, turning it into a full-fledged AI accelerator on the level of the Nvidia A100 80 GB, which costs about $20,000. At the same time, the modified card turned out to be even faster than the original flagship in some tasks.
A researcher under the nickname Kate was able to unlock the hidden hardware capabilities of the Nvidia CMP-170HX (8 GB) mining video card, turning it into a full-fledged AI accelerator on the level of the Nvidia A100 80 GB, which costs about $20,000. At the same time, the modified card turned out to be even faster than the original flagship in some tasks.
The CMP-170HX mining accelerators are based on the same GA100 GPU as the A100 server GPUs, Tom's Hardware reports . At the factory, Nvidia physically limits their capabilities using an electronic fuse controller (eFuse), disabling some of the shader units, the PCIe 4.0 bus, NVLink, and most of the HBM2e memory that is actually present on the chip substrate.
Previously, enthusiasts have already used the CMPUnlocker software mod, which allowed partially bypassing driver-level restrictions (opening up to 40–64 GB of memory and PCIe 2.0 mode). However, hardware fuses remained inaccessible for overwriting. Kate drew attention to an architectural flaw: in Nvidia's design, "overburning" additional fuses does not reduce functionality, but on the contrary - opens access to locked chip resources.
To force the write-protected eFuse controller to accept new settings, the researcher used a voltage failure attack method:
using the GPT-6 Astra model, she designed a simple printed circuit board based on the $8 STM32G4 microcontroller;
the board connected to the video card's power control lines (PWMVID) and interfered with the operation of the voltage regulators (VRM), creating short-term failures on the GPU power line;
By running a continuous cycle of reading the fuse status during power failures, the researcher put the controller into a state where the security bits allowed overwriting.
Since physical fuses cannot be rolled back, Kate carefully left those shader modules that were rejected at the factory as defective disabled, but unlocked all working resources.
The researcher applied the method to ten of her CMP-170HX graphics cards. Nine of them successfully used all 80 GB of video memory, and one - 64 GB. In addition, the unlocked cards received full support for PCIe 4.0, ECC error correction, NVLink interface and Multi-Instance GPU (MIG) mode.
Since the GA100 chip initially has 128 stream multiprocessors (SMs) with a factory-supplied margin (the standard A100 has 108 active), Kate cards managed to use more than 120 working units. Thanks to this, the performance in FP16 FMA calculations was 20% higher than that of the standard Nvidia A100.
On the secondary market, the 8GB and 10GB versions of the CMP-170HX currently cost around $2,500–3,000, while the full PCIe 80GB version of the A100 costs upwards of $20,000. Kate decided to publish her research after another enthusiast independently discovered the same vulnerability. She also alerted Nvidia, but the company will no longer be able to programmatically patch the hole on previously released graphics cards.
