A $3,000 Mining Card Turned Into an A100 Equivalent: From 8 to 80 GB of Memory
Running large neural networks at home takes VRAM, and a server-grade 80 GB NVIDIA A100 costs $20,000 or more on the secondary market. Researcher Kate took a $3000 CMP 170HX mining card, which the factory limited to just 8 GB, and bypassed the lock with a brief voltage dip at startup. As a result, 9 of 10 cards unlocked all 80 GB of memory, the same as the A100. The method is still a lab experiment: a single mistake can permanently kill the card.
AI-processed from TG: @neurogen_news (Neurogen); edited by Hamidun News
Running large language models locally instead of in the cloud takes VRAM, and VRAM is the most expensive part. The server-grade 80 GB NVIDIA A100 (the PCIe version) sells on the secondary market for $20,000 and up. A researcher who goes by Kate found a way to get nearly the same thing for $3000.
She took an NVIDIA CMP 170HX mining card, which the factory limited to just 8 GB of memory, and "unlocked" it to 64–80 GB. Imagine buying a laptop at half price, then finding out it has the same processor and all the memory inside, but the manufacturer forbade you from using them. That is roughly what happened here, except the lock had to be bypassed with a soldering iron.
The same chip, with a factory limiter
The CMP 170HX is built on the same GA100 die as the flagship A100. The difference lies in the settings. At the factory, NVIDIA "throttles" the chip with electronic fuses (eFuses): it disables most of the HBM memory (the fast memory mounted next to the chip) and many of the compute blocks. This lets the company sell the card to miners at a lower price without competing with its own server accelerators.
Software tools for removing the limits, such as cmpunlocker, have appeared before, but they unlocked only part of the capabilities. Kate went further and used a hardware method.
How the power glitch works
The method is called voltage fault injection. The idea is to slightly "sag" the power supply at a precise fraction of a second during card startup. At that moment the protection makes a mistake and lets through a write to the fuseless fuses area, which is normally locked against changes. After that, the factory limits are overridden in software, while the physical fuses remain untouched.
For this, Kate built a glitcher board around an STM32G4 microcontroller. According to the post's author, the model GPT-6 Astra helped design the board and firmware. The glitcher connects to the graphics card with two wires, and one small resistor has to be removed. The timing of the voltage dip cannot be caught by hand, so the glitch parameters are tuned automatically. This takes anywhere from an hour to two days per card.
What came out of it
According to the project's description, the results are as follows:
- Memory. Instead of the stated 8 GB, 64–80 GB of working HBM was unlocked. On 9 of 10 test cards, all 80 GB worked.
- Compute blocks. Instead of 70 SM blocks (the basic compute modules of a GPU), 120–128 were enabled. In a synthetic FP16 test, meaning half-precision calculations typical for neural networks, the card came out 1.2 times faster than a production A100. The reason is that NVIDIA disables some blocks in official A100s as a reserve against manufacturing defects.
- Interfaces. PCIe Gen4 and even NVLink were enabled, although this required a small modification to the board.
These figures come from the project description. The source mentions no independent replications, and a synthetic benchmark result does not guarantee the same speed in real training or model inference.
Risks and what to do in practice
Repeating this at home is dangerous. It takes careful soldering, a homemade glitcher, and patience. A faulty write to the fuses is irreversible, and it is easy to turn the card into a "brick." In addition, bypassing the manufacturer's protection may violate warranty and license terms.
Still, the story has practical takeaways. First, the line between NVIDIA's cheap and server hardware runs largely through settings and fuses, not through the silicon itself. Second, the secondary market for the CMP 170HX will most likely see a new wave of interest from AI enthusiasts: 80 GB of HBM for $3000 looks tempting next to an A100 at $20,000. Third, the method itself is telling: an AI assistant helped design a hardware board, and the barrier to entry for complex hacking projects is falling.
The research and source code are on GitHub in the ga100vfi repository. If you are building your own cluster for local models, the project is worth studying. But for now it is a lab experiment, not a way to cut hardware costs at scale for production.
- Nov 8, 2026SageMaker can now split one GPU into 7 parts: more models on the same hardware
- Nov 12, 2026RelayVSR upscales video by 4× and delivers 29 fps on an A100
- Nov 12, 2026Developer With No Coding Skills Vibe-Coded a Bedbug Map in 2 Weeks: 300K Users and Wars With Hotels
- Nov 12, 2026Father builds free photo-based carb counter: 3,000 users in 17 days
Key Questions Answered
What are the key details on The same chip, with a factory limiter?
The CMP 170HX is built on the same GA100 die as the flagship A100. The difference lies in the settings. At the factory, NVIDIA "throttles" the chip with electronic fuses (eFuses): it disables most of the HBM memory (the fast memory mounted next to the chip) and many of the compute blocks. This lets the company sell the card to miners at a lower price without competing with its own server accelerators.
How the power glitch works?
The method is called voltage fault injection. The idea is to slightly "sag" the power supply at a precise fraction of a second during card startup. At that moment the protection makes a mistake and lets through a write to the fuseless fuses area, which is normally locked against changes. After that, the factory limits are overridden in software, while the physical fuses remain untouched. For this, Kate built a glitcher board around an STM32G4 microcontroller.
What came out of it?
According to the project's description, the results are as follows: Memory. Instead of the stated 8 GB, 64–80 GB of working HBM was unlocked. On 9 of 10 test cards, all 80 GB worked. Compute blocks. Instead of 70 SM blocks (the basic compute modules of a GPU), 120–128 were enabled. In a synthetic FP16 test, meaning half-precision calculations typical for neural networks, the card came out 1.2 times faster than a production A100.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.