Nvidia CMP 170HX Voltage Mod: How Voltage Glitching Turns a Mining Card Into an 80GB A100

Nvidia disables parts of the GA100 chip on the CMP 170HX with eFuses, one-time fuse settings read when the GPU powers up. The voltage mod attacks that moment. Instead of changing the fuses, it makes the chip misread them.

According to Kate’s write-up, a small STM32G4 microcontroller board is wired to the card’s core voltage rail and its PWMVID control line, after removing a tiny 0-ohm resistor. The board then drives the card’s own MP2888B voltage regulator to produce precisely timed voltage dips while the fuse controller starts up. A comparator triggers at 225 mV, and the system sleeps and wakes the PC over and over until a glitch lands.

Finding a working glitch takes “somewhere between 1 hour and 48 hours” per card, and the right timing is specific to each chip. Kate says she published some of her tooling and disclosed the findings to Nvidia’s product security team before publishing.

The CMP 170HX launched in 2021 as part of Nvidia’s Cryptocurrency Mining Processor line, built to keep miners away from gaming GPUs. Under the heatsink sits the same GA100 die as the A100, but Nvidia crippled it for anything except mining: 8GB usable memory, 70 SMs, PCIe Gen 1 and no NVLink.

Kate’s research says the 8GB card physically carries about 80GB of HBM. An existing software tool, cmpunlocker, could already open up 64GB, but only the voltage glitch reaches the full 80GB and re-enables the disabled SMs.

That matters for miners: thousands of these cards were left over when Ethereum moved to proof of stake. If you are deciding what to do with old mining gear, see our guide to selling or renting hashrate as a small miner and how a small-scale crypto data center is built.

Kate tested ten 8GB cards: nine reached the full 80GB and one reached 64GB, roughly a 90% success rate. Unlocked dies typically showed 120–126 working SMs, more than the 108 SMs a retail A100 ships with.

In an OpenCL FP16 benchmark, a retail A100 scored 72.5 TFLOPS, a software-unlocked CMP 170HX 50.1 TFLOPS and the voltage-glitched card 85.9 TFLOPS, about 1.18 times the A100. The unlocked card also reported 22.8 TFLOPS FP32 and 11.3 TFLOPS FP64. Kate says LLM servers such as vLLM and SGLang run on it.

Hardware Busters, citing Tom’s Hardware, puts the mining card at roughly $2,500 versus about $20,000 for an A100. That gap is why AI hobbyists are watching. For what full-scale GPU capacity costs, see AI data center build costs.

This is not plug-and-play. Kate warns that glitch patterns are “very much die-specific”, and some glitch states leave a card “basically unusable”. The mod needs fine soldering, custom firmware and a dedicated test rig, and it voids any warranty.

• Hardware damage: wrong glitches can leave the fuse read in a broken state.
• NVLink: needs an extra board modification and host-side steps. Bridging cards on different power supplies connects their 3.3V rails.
• Results are claims: the numbers come from one researcher and have not yet been independently reproduced.
• Scam listings: expect “pre-unlocked 80GB A100” listings. Check sellers with our scam website checker.