Expert Turns Nvidia’s 8 GB Mining Card Into an 80 GB Accelerator

An expert named Kate used a voltage glitching technique to turn an 8 GB Nvidia CMP-170HX mining card into an accelerator with 80 GB of memory. The modified card also outperformed the more expensive Nvidia A100 in some computations. The 8 GB version currently sells for about $3,000 (5,100 manats). The 10 GB version costs $2,500 (4,250 manats). An A100 with 80 GB of memory can cost $20,000 (34,000 manats).

The modification works because the memory blocks remain physically present on the GA100 chip. In lower-end models, the manufacturer disables part of the high-speed HBM2e memory and some compute modules using electronic fuses. The card’s 16,384 fuses control various components, including the memory and compute modules. They also control PCIe 4.0 connectivity, the NVLink system that provides a high-speed connection between cards, and video hardware. The 8 GB version is actually equipped with chips corresponding to 80 GB of memory, while the 10 GB version has chips corresponding to 40 GB.

Kate developed a separate circuit board to bypass the protection of the system that controls the fuses. Kate used the GPT-6 Astra model to design the circuit. The board’s main component is an $8 (about 14 manats) STM32G4 microcontroller. The device interfered with the PWMVID signals that regulate the card’s voltage, creating brief voltage spikes. As a result, the fuse controller began accepting modification commands.

During testing, 80 GB of memory worked on 9 cards, while 64 GB of memory was used on another card. After the functional compute modules were enabled, the card’s FP16 FMA performance in 16-bit computations was 20% higher than that of the standard A100. Kate attributes this to the chips retaining 120 or more functional modules instead of 108. The GA100 chip is manufactured with a total of 128 modules to reduce the impact of manufacturing defects.

The previous CMPUnlocker software modification could unlock only 40–64 GB of memory. In addition, the card’s connection speed was limited to PCIe 2.0. ECC, which corrects memory errors, NVLink, and the ability to run multiple virtual GPUs also could not be enabled. The new voltage glitching technique eliminated these limitations. The most successful test card operated at PCIe 4.0 speed with all server features enabled.

The process is irreversible. Therefore, re-enabling modules deemed defective at the factory could render the card unusable. Kate kept the risky modules disabled and did not test the video encoding section. The research was published after another person independently discovered the same vulnerability. Nvidia was notified of the issue, but preventing this technique on cards that have already been sold appears difficult.