Is it true that a mining GPU was "turned into an 80 GB A100" with a cheap homemade board?

Partly true. The CMP 170HX is a 2021 cryptocurrency mining card built on the same GA100 chip as the A100 server accelerator, but with cut-down capabilities: 8 GB of memory, 70 compute units and a slow connection to the computer (4 lanes of first-generation PCIe). Kate writes that after the hack the card becomes "effectively an A100-80G": more compute units (about 120–126), the full 80 GB of memory and a fast PCIe 4.0 connection. But the system sees it not as an A100 but as an "NVIDIA Graphics Device"; it needs "a simple driver patch" (according to Tom's Hardware, the drivers still require the CMPUnlocker software unlock); linking several cards (NVLink) requires another board modification; and she did not test some features (ECC, MIG, video decoders).^{[1][4][5][7][8]}

True. Kate writes that she asked the GPT-6 Astra model to design a board based on an STM32G4 microcontroller that connects to the card's power management circuit; Tom's Hardware says the same. The schematic, firmware and software are published on GitHub.^[1][5][2]

Misleading. Kate's write-up gives no price at all. Tom's Hardware writes that the board is built on an $8 STM32G4 microcontroller: that is the price of the chip itself, not of a finished board, let alone the whole hack. Besides the board, you need to desolder a resistor on the card, solder wires to the power stage, and spend 1 to 48 hours tuning the glitch for each card.^{[1][5][10][7]}

The original source has no prices; they come from the media. Tom's Hardware writes that the 8 GB CMP 170HX now costs about $3,000, the 10 GB version about $2,500 (the headline of the same article says "$2,500" for 8 GB), and the 80 GB PCIe A100 costs $20,000 or more; the outlet warns that prices will swing a lot as word of the mods spreads. The card's price has already jumped: according to iXBT on August 8, 2026, used CMP 170HX cards used to sell for $100–200, and after a software memory unlock appeared, for $1,500–2,200.^[5][7][9]

Partly true. The memory does not come out of nowhere: the chip physically has 80 GB, but the factory enabled only 8 GB. Kate unlocked the full 80 GB on cards where all of that memory works. For the 10 GB version of the card, she says, very few units reach 80 GB, and they run unstably. There is no "infinite memory" as the post promises.^[1][5][8]

Misleading. This refers to Kate's own 10 cards of the 8 GB version: nine had all 80 GB working, one had only 64 GB (Tom's Hardware reports the same). She herself calls the sample limited. It is the share of cards with working memory, not the success rate for everyone who tries: nobody has independently reproduced the method yet, and tuning the glitch, she says, depends heavily on the specific chip.^[1][5][8][7]

True for one test. In ProjectPhysX's synthetic half-precision (FP16) benchmark, the modified card scored 85.86 TFLOP/s versus 72.49 for a stock A100, 1.18 times more; Tom's Hardware rounds it to 20%. This is because the modified card has more compute units enabled (often 120 or more) than the 108 on the A100 on sale. This is the researcher's own measurement; she says AI software (vLLM, SGLang) works, but gives no comparisons for it.^[1][5][10]

The original source says nothing like this. Kate only writes that factory settings decide whether a GA100 chip becomes a "mining card" or an "AI datacenter workhorse." She is talking about one card, not a data center.^[1]

False. Kate's write-up describes vulnerability research, not a step-by-step guide. The GitHub code is a set of materials for experiments (board schematic, firmware, control software), with no beginner instructions and no risk warnings.^[1][2]

No (the post implies this without saying it outright). According to the original source and retellings: you need to solder on the card's board; after a failed glitch the card may boot "practically unusable"; regular fuses (tiny links inside the chip that store factory settings) cannot be restored once blown (Tom's Hardware: there is no "undo" button; only some settings stored separately can be reset); tuning the glitch takes up to 48 hours; and linking cards via NVLink when they run on different power supplies risks shorting the 3.3 V power rails.^{[1][5][8][7][10]}

CMP 170HX specs described: GA100 chip, 8 GB of HBM2e memory, connection limited to 4 lanes of first-generation PCIe, a card for mining farms.^[4]

A software unlock (CMPUnlocker) opens up to 64 GB of memory on the 8 GB CMP 170HX; card prices rise roughly tenfold.^[9][6]

Kate publishes her research: a power glitch at card startup lets her rewrite the chip's factory settings and unlock 80 GB and more compute units. The code is on GitHub; NVIDIA was notified before publication.^[1][2]

Tom's Hardware, Hardware Busters and others retell the research with prices of "$2,500–3,000 → $20,000" and an "$8" figure for the microcontroller.^[5][7]

A translation of the write-up on Habr (a Russian tech site) and a retelling in a user blog on Overclockers.ru: "Enthusiast uses AI… beats the original by 18%."^[8][10]

A post, "Woman techie uses AI to turn…," with prices for the card, the A100 and the board, "9 out of 10," "by 18%," "run a data center" and a call to "grab the guide to infinite memory"; it links to Kate's write-up.^[3]

Research write-up: power glitch, bypassing fuse protection, 80 GB on cards with working memory, 1.18× an A100 in the FP16 test, NVLink risks, regular fuses are irreversible.^[1]

Detailed retelling: a board on an $8 microcontroller designed by GPT-6 Astra; 9 cards with 80 GB and one with 64 GB; FP16 20% above an A100; prices of about $2,500–3,000 for the CMP 170HX and $20,000+ for the A100, which "will fluctuate wildly"; NVIDIA was notified but can do little about cards already sold.^[5]

Stresses that this is not "plug and play": wires to test points, a desoldered resistor, up to two days of tuning per card; no independent reproduction.^[7]

Russian translation of Kate's write-up, including the caveats about risks and dependence on the specific chip.^[8]

Retelling that emphasizes "with AI" and "by 18%"; mentions "8 dollars" citing Tom's Hardware; notes irreversibility and soldering.^[10]

The post through which the news reached the site owner: a retelling of the research with inflated conclusions and a call to "grab the guide."^[3]

2026-11-12 · 28.3K views as of 2026-11-12

Specialists are interested in the code; the page shows no reports of anyone reproducing the hack.^[2]

2026-11-12 · 80 stars and 37 forks as of 2026-11-12

Hasil: Partly true: the CMPUnlocker program unlocked up to 64 GB of memory, but the connection only up to second-generation PCIe; prices for these cards rose roughly tenfold after the news.^[9][6][1]