Voltage Glitch Unlocks 80GB From Nvidia CMP-170HX Card

By Grace Kim

An $8 voltage-glitch circuit enabled nine Nvidia CMP-170HX mining cards to expose 80GB of HBM2e memory, matching a PCIe A100 whose 80GB configuration costs $20,000 or more. The result came from hobbyist testing, not an Nvidia product or benchmark program. The work targets local-AI builders using used data-center GPUs instead of DGX Spark or Apple Mac Studio systems.

How did the 8GB mining card reach 80GB?

Kate’s route began with a paper describing an exploit of GA100 microcode. It targeted the normally write-protected onboard eFuse controller. Kate said: “Nvidia’s design is insecure, as popping fuses can grant access to more features, not fewer.”

Factories use these physical fuses to disconnect parts that failed testing, allowing a lower-end product to be sold without wasting an entire die. The GA100 silicon already contains the corresponding HBM2e memory dies and 128 streaming multiprocessors, or SMs. Factory fuses disable selected memory sections and shader modules.

The 80GB CMP-170HX derives from the 80GB GA100 configuration. The 10GB card derives from the 40GB version. That architecture explains how cards can expose only 8GB or 10GB while their dies contain 40GB or 80GB of memory. The disabled memory remains physically present. The task was to change the fuse configuration without damaging the silicon.

How could a read-only fuse controller be changed?

The Nvidia controller contains 16,384 eFuses in a 512-by-32 array. Readable status fields cover memory sections, individual SMs, PCIe 4.0, NVLink and video-codec hardware. Under normal conditions, software can read those fields but cannot write new values.

The board, designed with GPT-6 Astra, used an $8 STM32G4 microcontroller to tap power signals on the card. The card’s voltage regulators send PWM signals under Nvidia’s PWMVID standard. The board intercepted those signals and forced voltage glitches onto the GPU power line near the fuse controller.

Repeated fuse-status reads during the glitches moved the controller into a state where its security flags permitted writes. Reconfiguration then became straightforward. Kate forced previously defective shader modules to remain disabled, because those physical fuse decisions cannot be undone.

What did the modified cards deliver?

Kate applied the method and tested multiple workloads. Nine cards exposed full 80GB of working memory. A tenth card “only” had 64GB usable. Her test also found FP16 FMA performance 20% above a regular A100 after she unlocked all shader modules.

She attributed the gain to chips retaining 120 or more usable SMs, compared with 108 on the A100 specification. Each GA100 die starts with 128 SMs, leaving enough working units to route around manufacturing defects. The findings cover Kate’s FP16 FMA tests; they do not establish faster performance in every workload.

What did the voltage mod add over CMPUnlocker?

CMPUnlocker, a software modification for CMP-170HX cards, remains required to activate the newly exposed functions in drivers. It does not remove four limits:

• VRAM remains capped at 40GB or 64GB, depending on the card variant.
• Many shader modules remain locked.
• PCIe bandwidth remains at 2.0.
• ECC, NVLink and Nvidia multi-instance GPU support remain unavailable.

Those omissions matter in server deployments and preserve Nvidia’s product segmentation. Kate’s voltage modification removed those specific limits. Her best card ran at PCIe 4.0 with ECC, NVLink and multi-instance GPU support restored. She did not test the video-encoder blocks, leaving their operation unverified.

What do the modified cards cost?

At publication, an 8GB CMP-170HX hovered around $3,000, while the 10GB version cost about $2,500. A PCIe A100 with 80GB reached $20,000 or more. SXM server-module versions cost less, but require surrounding infrastructure. The key comparison is a $3,000 mining card against at least $20,000 for the A100.

Used CMP prices will likely move as more enthusiasts learn about the mods. The circuit board itself cost $8, while the starting card cost thousands of dollars.

What happens next?

Kate said she published the research after another person independently found the same issue. She notified Nvidia of the vulnerability. Existing cards present limited options for a fix because the modification changes physical fuse states. Hardware can therefore carry capabilities its factory configuration was designed to withhold.

As hobbyists reproduce the voltage exploit, CMP-170HX resale prices and Nvidia’s ability to contain the flaw in future products will shape the commercial response.