A researcher under the nickname Kate was able to unlock the hidden hardware capabilities of the Nvidia CMP-170HX mining graphics card (8 GB), turning it into a full-fledged AI accelerator on par with the Nvidia A100 at 80 GB, which costs about $20,000. At the same time, in some tasks the modified card turned out to be even faster than the original flagship.

What the vulnerability is about

Tom’s Hardware writes that CMP-170HX mining accelerators are built on the same GA100 graphics processor as the A100 server chips. At the Nvidia factory, their capabilities are physically limited 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 had already used the CMPUnlocker software mod, which allowed partially bypassing driver-level restrictions (unlocking up to 40–64 GB of memory and PCIe 2.0 mode). However, the hardware fuses remained unavailable for rewriting. Kate drew attention to an architectural flaw: in Nvidia's design, "reblowing" additional fuses does not cut functionality, but on the contrary — opens access to the chip's locked resources.

How the volt mode worked

To force the write-protected eFuse controller to accept new settings, the researcher used a voltage fault injection attack method:

  • using the GPT-6 Astra model, she designed a simple printed circuit board based on an STM32G4 microcontroller costing $8;

  • the board was connected to the graphics card's power control lines (PWMVID) and interfered with the operation of the voltage regulators (VRM), creating short-term failures on the graphics processor's power line;

  • by running a continuous loop of reading the fuse status during voltage failures, the researcher put the controller into a state where the security bits allowed rewriting.

    • Since physical fuses cannot be rolled back, Kate carefully left disabled those shader modules that had been rejected at the factory as defective, but unlocked all functional resources.

    Results and consequences

    The researcher applied the method to ten of her CMP-170HX graphics cards. In nine of them, all 80 GB of video memory successfully worked, and in one — 64 GB. In addition, the unlocked cards received full support for PCIe 4.0, ECC error correction, the NVLink interface, and Multi-Instance GPU (MIG) mode.

    Since the GA100 crystal initially has 128 streaming multiprocessors (SMs) with a margin for factory defects (in the standard A100, 108 are active), the Kate cards managed to utilize more than 120 working blocks. Thanks to this, performance in FP16 FMA calculations turned out to be 20% higher than that of the standard Nvidia A100.

    On the secondary market, versions of the CMP-170HX with 8 GB and 10 GB currently cost about $2,500–3,000, while the price of a full-fledged PCIe version of the A100 with 80 GB reaches $20,000. Kate decided to publish her research after another enthusiast independently discovered the same vulnerability. She also warned Nvidia, but the company will no longer be able to close this hole in software on previously released graphics cards.