NVIDIA sold its CMP 170HX mining card to mining farms with only 8GB of memory and 70 compute units enabled. But the chip inside is the same one used in the A100 data center GPU. A hardware security researcher named Kate published her findings on November 10. She used voltage fault injection to trick the fuse controller that locks features on the chip. Most of the cards she worked on now run with the full 80GB of memory and 120 to 128 compute units (SMs). Her half-precision (FP16) speed came out about 18% higher than the retail A100.
How the lock works, and how it was opened
One chip can be sold in different tiers because of fuses built into it. At the factory, some of those fuses are blown. That permanently turns off the matching memory, compute units and interfaces, and software cannot switch them back on. Earlier, the community had found software-only unlocks. Those got memory up to 64GB at most, added no compute units, and kept PCIe at 2.0.
This time, she went around the rule that blowing fuses has to go through a secure process. At the moment the chip powers on, she creates a very short disturbance in the supply voltage. The fuse controller briefly enters an abnormal state. In that window, new settings can be written, and the disabled features come back on. She built her own control board to produce this disturbance. It connects to the card's power control signal. If an attempt fails, she cuts the power and tries again.
What the unlocked card can do
Here are the author's test results.
- Memory: Of the ten 8GB cards she had, nine unlocked to a usable 80GB. One only reached 64GB.
- Compute units: They went up from 70 to somewhere between 120 and 128. The exact number depends on each chip's quality. The retail A100 has 108.
- Compute speed: Half-precision (FP16) measured about 85.9 TFLOPS. The retail A100 scores 72.5. The software-only unlocked card reaches only 50.1.
- Interface: PCIe can be restored to 4.0, which needs one extra circuit change. NVLink, used to link multiple cards, can also be enabled, but that takes hardware changes. The author says the details will come in a later post.
- Real workloads: The 80GB of memory passed several rounds of read and write checks. Large model inference frameworks like vLLM and SGLang run normally.
Why this matters
For people who want to run large models locally, 80GB of memory with near-A100 compute is exactly the part that holds back large model work. The research got more than 300 points on Hacker News, and the discussion has been lively. The author did not list a price for the cards, so check used market prices yourself.
For hardware security, it shows that using fuses to separate product tiers is not reliable. If you can change the voltage at power-on, features that were "permanently" disabled can come back. The author says she told NVIDIA's product security team before publishing.
Don't rush to buy a card
- High bar: You would need to solder a control board yourself, fly wires on the card's circuit board, and remove components. This is hardware security research work, not something you can install with software.
- Time-consuming: The author says each card took 1 to 48 hours of repeated tries to succeed.
- May destroy the card: Many failed attempts leave the fuse controller unresponsive, and the card is basically ruined. Circuit changes are permanent, and the warranty is gone too.
- Results vary by card: The same method works differently on different chips. Most 10GB cards cannot reach 80GB.
- Limited data: Only 10 cards were tested. There is no data yet on long-term stability or cooling.
- Modifying and reselling modified hardware may break vendor terms or local laws. Check those before you act.
Comments (0)
No comments yet. Start the conversation.