# Voltage Fault Injection Unlocks Full GA100 Performance on CMP-170HX Mining Cards

Researchers bypass factory-set hardware restrictions to transform budget mining GPUs into high-performance A100 units.

- **Source:** [Hacker News](https://news.ycombinator.com)
- **Published:** 2026-11-11
- **Topic:** [technology](https://brifing.si/c/technology)
- **Original article:** <https://blog.kate.land/posts/using-voltage-glitching-to-unlock-the-a100/>
- **On Brifing:** https://brifing.si/s/12297a0d-e503-41d2-9156-777ab37eed01

## Hardware Segmentation Vulnerability

A security researcher identified a critical voltage fault injection vulnerability within the GA100 silicon die. By manipulating the onboard fuse controller, an operator can override factory-set product segmentation. This process effectively transforms restricted CMP-170HX mining cards into fully functional, high-end A100-80G units, bypassing the original hardware limitations imposed by the manufacturer.

## The Role of eFuses

Nvidia utilizes an elaborate eFuse controller, containing 16,384 fuses in a 512x32 array, to segment products. These fuses physically disable shader modules, VRAM stacks, and features like NVLink or PCIe Gen4. Typically, these fuses are set during manufacturing to account for yield defects, ensuring that only stable, lower-tier chips reach the market.

## Voltage Glitching Mechanism

The modification involves using an $8 STM32G4 microcontroller to intercept PWM signals sent to the GPU's voltage regulators. By inducing precise voltage glitches, the researcher forces the fuse controller into an insecure state. This allows the user to write to registers that are normally read-only, effectively reconfiguring the card's internal hardware status.

## Performance Gains Achieved

Unlocked cards demonstrate significant performance improvements, achieving FP16 FMA throughput approximately 1.2 times faster than a retail A100. This is possible because the GA100 die contains 128 Streaming Multiprocessors, whereas retail A100 units are limited to 108. Enabling these extra functional units results in superior raw computational power compared to official enterprise specifications.

## Memory Capacity Expansion

The GA100 silicon often contains more HBM2e VRAM than is enabled in lower-tier SKUs. By manipulating the fuse controller, researchers have successfully unlocked up to 80GB of VRAM on cards originally sold with 8GB or 10GB. While results vary based on the specific die, most test units successfully accessed the full 80GB capacity.

## Restoring Enterprise Features

Beyond raw compute, the modification restores critical enterprise features previously disabled by Nvidia. These include PCIe Gen4 bandwidth, NVLink, and Multi-Instance GPU support. While the researcher has not yet verified NVJPG or NVDEC video codec functionality, the hardware-level unlock suggests these features are theoretically accessible through the same manipulation of the fuse array.

## Cost-Benefit Comparison

A retail Nvidia A100 80GB PCIe card typically costs around $20,000. In contrast, a CMP-170HX mining card is available for approximately $2,500 to $3,000. This modification offers a massive cost reduction for enthusiasts building local AI clusters, providing enterprise-grade hardware performance at a fraction of the official market price for data center equipment.

## Risk of Physical Damage

The modification carries inherent risks, as it involves permanent changes to physical fuses. If a user accidentally enables shader modules that were originally disabled due to manufacturing defects, the card may become unstable or fail entirely. There is no 'undo' button for these physical hardware changes, making the process a high-stakes operation for users.

## Research Disclosure Process

The researcher decided to publish these findings after learning that another individual had independently discovered the same vulnerability. The details were disclosed to Nvidia's product security team prior to public release. The researcher believes that making this information available is in the public interest, given the widespread nature of the hardware vulnerability.

## Limitations of Soft-Mods

Previous 'CMPUnlocker' soft-mods were limited to driver-level changes and could not bypass physical hardware restrictions. These earlier methods failed to enable ECC memory, NVLink, or full PCIe 4.0 speeds. The new voltage-injection method is distinct because it modifies the physical state of the chip, overcoming the limitations that software-only approaches could not address.

## Industry Security Implications

The vulnerability highlights a fundamental flaw in the design of the GA100 fuse controller. In a secure system, blowing a fuse should only induce a more restrictive state that cannot be reversed. Nvidia's design, however, allows for the relaxation of restrictions by manipulating the fuse array, which is a significant departure from standard hardware security practices.

## Future Research Directions

The researcher plans to continue investigating the GA100 die and its fuse controller. Future efforts may focus on testing the remaining disabled features like video encoding and refining the voltage injection process for better reliability. As awareness of these mods grows, the market for used mining GPUs may see significant shifts in demand and pricing.

## Sources covering this story

- [Hacker News — Unlocking NVIDIA A100 Performance via Voltage Fault Injection](https://blog.kate.land/posts/using-voltage-glitching-to-unlock-the-a100/)
- [Tom's Hardware — Techie Hacks Nvidia CMP-170HX to Match $20,000 A100 Performance](https://www.tomshardware.com/pc-components/gpus/techie-turns-usd2-500-cmp-170hx-8gb-mining-card-into-a-full-fat-a100-80gb-with-voltage-glitching-carefully-blowing-onboard-fuses-to-unlock-entire-vram-and-shader-modules)

---

_Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._
