Research

Nvidia CMP 170HX to A100 80GB: What the Voltage Mod Does

A researcher used voltage glitching to turn an 8GB Nvidia CMP 170HX mining card into an 80GB A100-class GPU. The published numbers, the limits, and what it means for local LLMs.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: November 2026

On November 10, 2026 a hardware researcher who blogs as Kate published a method that turns Nvidia's CMP 170HX mining card into what she calls "effectively an A100-80G". The 8GB card uses the same GA100 chip as the A100, with most of its memory and compute locked by fuses. The mod uses voltage glitching to change those fuses. It is irreversible, takes between 1 and 48 hours per card, and has no published LLM speed numbers yet.

What Changes

ItemStock CMP 170HXAfter the hardware modRetail A100
Usable memory8 GB80 GB on cards with intact memory80 GB
Streaming multiprocessors enabled70About 120 to 126108, per Hardware Busters
FP16 throughput, OpenCL benchmark50.1 TFLOP/s with the software unlock onlyAbout 85.9 TFLOP/s72.5 TFLOP/s

On that one benchmark the modified card is about 1.2 times a retail A100 and 1.7 times the software-only unlock. These are the author's measurements on the ProjectPhysX OpenCL benchmark.

How the Mod Works

The attack targets the fuse controller on the GA100 chip. A custom board intercepts the signal that sets the card's core voltage and briefly disturbs it at the right moment, which lets the operator blow fuses the chip would normally protect. Hardware Busters reports the board is built on an STM32G4 microcontroller. The author says she disclosed the issue to Nvidia's product security team before publishing.

Limits

  • It cannot be undone. Fuse changes are permanent.
  • Results vary by card. Of 10 cards examined, nine reached about 80 GB and one reached 64 GB. The 10GB version of the card does not reliably reach 80 GB.
  • It is slow. Glitching takes "between 1 hour and 48 hours" per card.
  • Some features need more work. NVLink and PCIe Gen4 x16 need further hardware changes. Video decode, JPEG decode and multi-instance GPU were not tested.
  • It needs custom hardware and soldering. This is not a firmware flash.

Does It Matter for Local LLMs?

The appeal is 80 GB of memory on one card, which is enough to hold models that otherwise need two or three consumer GPUs. The author writes that real workloads, naming vLLM, are fully functional. She published no tokens-per-second figures, so there is no measured LLM speed to compare with a RTX 5090 or a Mac Studio. The A100 is a 2020 design, so expect it to trail current cards on speed while matching or beating them on memory.

What the Card Costs

The price is unclear. The author's post gives none. One Tom's Hardware headline on this story calls it a $2,500 card, and an earlier Tom's Hardware headline on the software unlock called it a $250 card. We could not confirm either figure, and used prices are likely to move now that the method is public.

Should You Try It?

  • If you want a working local setup: no. Buy hardware with a warranty. See local LLM hardware under $5,000.
  • If you do hardware research: the write-up is detailed, and a second researcher reportedly found the same weakness.
  • If you already own these cards: wait for independent replications and LLM benchmarks before modifying one.

Methodology

Facts come from the author's post of November 10, 2026, with Hardware Busters for the retail A100 comparison. Presenc AI did not test the mod and no independent benchmark had been published by November 13, 2026.

Frequently Asked Questions

According to the researcher who published the method on November 10, 2026, yes on most cards. Nine of the 10 cards she examined reached about 80 GB of usable memory and one reached 64 GB. No independent benchmark had confirmed it by November 13.
No LLM speed figures have been published. The only numbers are from an OpenCL compute benchmark, where the modified card reached about 85.9 TFLOP/s at FP16 against 72.5 for a retail A100.
No. It works by permanently changing fuses on the chip, and the process takes between 1 and 48 hours per card with custom hardware.
CMPunlocker is an earlier software unlock. The hardware mod builds on it and goes further: on the published benchmark it was about 1.7 times faster than the software unlock alone, and it restores the full 80 GB of memory.

Hire agents, not agencies.

Presenc is a team of AI agents that builds the pages your buyers search for, and a person checks every one. We take on a few teams at a time.

Request access