Hobbyist unlocks 80GB on Nvidia's crippled mining card with a $5 Raspberry Pi Pico
A voltage-glitching method published on 10 November turns Nvidia's 8GB CMP 170HX mining card into an A100-class 80GB GPU; a hobbyist reports doing it with a Raspberry Pi Pico.
Confirmed
- Voltage-glitching method published 10 November unlocks full 80GB on Nvidia's CMP 170HX mining card
- CMP 170HX is built on the same GA100 chip as Nvidia's A100
- Nvidia's CMP 170HX mining card is an 8GB card before the method is applied
Not yet known
- Unconfirmed: whether a $5 Raspberry Pi Pico can repeat the voltage-glitching method
- Whether the Raspberry Pi Pico repeat claim can be independently verified
What we know
A hobbyist has turned an 8GB Nvidia mining card into an 80GB AI accelerator using a Raspberry Pi Pico, a few hair-thin wires and a technique called voltage glitching. The method was published by a security researcher on 10 November. Five days later, a user on Reddit's LocalLLaMA forum posted a screenshot of a CMP 170HX reporting 81,920MiB of memory.
The CMP 170HX was sold during the crypto boom as a mining-only card. Under its heatsink sits the same GA100 chip as Nvidia's A100, the data-center GPU that trained a generation of AI models. The 8GB version carries about 80GB of HBM memory on the package, according to the researcher's write-up, but only 8GB is switched on. The card also runs 70 compute units (SMs), where a retail A100 runs 108, and its PCIe link is slowed.
The locks are set with eFuses, tiny one-time switches inside the chip that Nvidia blows at the factory. Chipmakers use them to sell the same silicon at different prices and to switch off parts that failed testing. HBM stacks are bonded to the chip and cannot be removed, so a GA100 with faulty memory could still be sold as a mining card. Ethereum mining needed only about 4GB when the card launched; the cap would have made it obsolete for mining around March 2027, had Ethereum not moved to proof-of-stake in 2022.
How the glitch works. Normally a blown fuse can only make a chip more restricted. But some of Nvidia's fuse settings work by writing new rows that override older ones, so a later write can lift a restriction. A safety setting blocks those writes on production cards. The attack removes one tiny resistor between the GPU and its power controller, watches the core voltage and, about one second into boot, sends three dips of roughly 2 milliseconds each. When the timing lands, the safety setting reads as off and new fuse rows can switch the disabled memory and compute units back on.
It is trial and error: a successful unlock took between 1 and 48 hours of repeated attempts, depending on the chip. The fuse writes are permanent, and each card needs its own, matched to that chip's real defects. The researcher says they disclosed the flaw to Nvidia's product security team before publishing, and that another researcher found it independently.
The results were large. Nine of ten 8GB cards gave the full 80GB, tested with memory checks and the vLLM and SGLang inference servers. Working compute units rose from 70 to about 120-126, and an FP16 benchmark reached 85.9 TFLOP/s, 18% above a retail A100's 72.5. With extra board fixes the cards also ran PCIe Gen4 x16 and NVLink. The 10GB version of the card is different: it carries about 40GB, and claims of 80GB on it are largely caching artefacts, the post says.
Fifteen hours with a Pico. The Reddit user swapped the researcher's custom glitch board for a Raspberry Pi Pico 2, a microcontroller that sells for about $5. Photos show enamelled copper wires running from the Pico to small parts near the card's power circuitry. The user said they finished after nearly 15 hours of work, had an AI assistant write the script for the fuse-writing step, and plan to unlock a second card. Their screenshot, dated 15 November, shows Nvidia's nvidia-smi tool listing the card with 81,920MiB, the size of a full A100 80GB. A setup service on the same screen reports a failure, so the post does not yet show the card running a model. Global Eye 24 has not verified the result.
"An AI wrote the script that unlocked more VRAM to run more AI. The loop is complete," read the top reply, with close to 300 upvotes.
Why it matters. Memory is the bottleneck for running large AI models at home, and 80GB on one card is rare outside the data center. A working unlock would turn surplus mining hardware into A100-class machines for local AI; one commenter said the cards already sell for close to their original price of over $4,000. The risks are real: the fuse writes cannot be undone, a wrong write can leave a card useless, and the work needs fine soldering on a board worth thousands of dollars. About one card in ten in the researcher's sample had too much faulty memory to gain anything. Nvidia has not commented.
By the Global Eye 24 desk, with AI assistance. Updated as the story develops. Spot an error? Write to corrections@globaleye24.com.