Tinkerer Unlocks 80GB on a $2,500 Nvidia Mining Card by Glitching Its Fuses

A hobbyist who goes by Kate has found a way to turn Nvidia's cheap CMP-170HX mining cards into something close to a $20,000 data center GPU, Tom's Hardware reported. By using voltage glitches to get around the cards' protected fuse controller, she unlocked up to 80GB of video memory and the full set of chip features on cards that sell for around $2,500 to $3,000.

Kate has told Nvidia about the vulnerability, though Tom's Hardware and Mezha.net both note the company has limited options for cards already on the market.

The key numbers

80GB of memory: Nine of Kate's test cards ran with the full 80GB of VRAM after the mod, while one had 64GB usable.
20% more performance: Unlocking all the shader modules gave her 20% more FP16 FMA performance than a regular A100, in her tests.
$2,500 vs. $20,000: The 10GB card costs around $2,500 and the 8GB version around $3,000, while a PCIe A100 80GB can cost $20,000 or more.
$8 microcontroller: Kate's custom circuit board is built around an STM32G4 chip that costs $8.
16,384 fuses: The cards carry that many eFuses, arranged in a 512 x 32 array.

How the hack works

Each card is built on Nvidia's GA100 chip, which Tom's Hardware says already contains all the memory dies, including those switched off at the factory. The 8GB card comes from a GA100 variant with 80GB of HBM2e memory, while the 10GB cards come from the 40GB variant. Factories normally use onboard fuses to disable parts of a chip that did not pass testing so the chip can be sold in a lower-end product.

Users cannot normally write to those fuses. Kate started from a paper describing an exploit in GA100 microcode and focused on the fuse controller. She had GPT-6 Astra design a circuit board that taps the card's power signals and intercepts the signals sent to the voltage-regulator controller, creating voltage glitches on the line that feeds the GPU. While the glitches ran, she read fuse statuses in a tight loop until the controller's security bits allowed writes.

She disabled any shader modules already marked as defective first, because the fuses are physical and cannot be reset. Kate said Nvidia's design is insecure because changing fuses can enable more features rather than restrict them.

Why it matters: what the hack changes for enthusiasts

Tom's Hardware says used data center GPUs are popular with enthusiasts building local AI clusters. A software mod called CMPunlocker already unlocks some abilities on these cards, but it is still limited: VRAM tops out at 40-64GB depending on the variant, many shader modules stay locked, PCIe bandwidth is limited to 2.0, and there is no ECC, NVLink or multi-instance GPU support. Those features are used in server setups and for product segmentation.

Kate's modification removes those limits. Her best test card ran at PCIe 4.0 speed with the full feature set. She did not test the video-encoder hardware. She attributed the extra performance to the chips often having 120 or more usable shader modules, compared with 108 on the spec sheet; a GA100 die starts with 128 to allow for manufacturing defects.

The reports caution that prices may change significantly as more people learn about the mods, and that the method involves permanently altering the hardware.