A hardware researcher who publishes as Kate says she has unlocked memory and compute that Nvidia switched off at the factory on its 170HX mining card, using voltage fault injection to burn normally locked fuses during boot

1

2

. Because the bypass operates at the hardware level, it goes further than the driver patches an earlier unlock project relied on

1

2

.

Source: Wccftech

Source: Wccftech

Per the researcher's own testing, the unlocked cards report the full 80GB of high-bandwidth memory physically present on the board, along with 120 to 126 working streaming multiprocessors

1

2

. That is more than the 108 Nvidia enables on the retail card it sells to data centers

1

2

. The 170HX arrived in 2021 as a passively cooled card for Ethereum mining

1

. It carries the same underlying chip as Nvidia's data-center accelerators but shipped with most of its memory and compute disabled

1

. Nvidia marketed the split under a slogan that cast GeForce as the gaming chip and the mining line as the mining one

1

. What makes the hidden hardware worth prying open now is the memory market. By one outlet's framing, high-bandwidth memory is effectively sold out amid the shortage

1

.

One benchmark, many gaps

In a single ProjectPhysX benchmark of half-precision, non-tensor math, the modified card ran about 18% faster than Nvidia's retail data-center accelerator

1

2

. That is the whole basis of the speed claim, and it is a narrow one. The test measures non-tensor throughput only. Tensor performance, error correction, memory partitioning and the video engines are untested, NVLink is not fully working, and a faster PCIe generation needs further hardware work

1

2

. Without those pieces, the card is not a demonstrated replacement for data-center work

1

2

.

Source: TechRadar

Source: TechRadar

The unlock works, per Kate's write-up, because the chip's fuse scheme runs the wrong way. A secure design should let fuses make a chip more restricted only, but this one lets an operator, as she puts it, "relax restrictions by blowing more fuses"

1

. The flaw, in her telling, sits in shipped silicon that Nvidia cannot fix remotely

1

. An earlier community project, cmpunlocker, had already used patched drivers to expose 64GB on the 8GB card, but it stopped at an older PCIe generation

1

2

. The attack technique itself is established in hardware security. The same class of voltage glitch has been used against a rival chipmaker's secure processor and to jailbreak Tesla's in-car computers

1

. Kate reports that OpenAI's GPT-6 Astra helped design the glitching board

1

.

Disclosure, and no comment

Kate notified Nvidia's product security team before publishing, per her write-up, and went public only after a second researcher independently reached the same finding

1

. Nvidia has not commented

1

. The finding sits within a recent run of Ampere-era Nvidia security research, from academic Rowhammer attacks on the company's GPUs to Jon Pry's June 2026 paper on the GPU security coprocessor

1

.

Slow, permanent, and not for everyone

One of ten tested 8GB cards topped out at 64GB of healthy memory

1

2

. Secondhand 170HX prices have climbed from roughly $100-250 before the first software unlock to over $1,000 since

1

2

. The data-center accelerator, for comparison, trades near $20,000, though both figures come from volatile resale listings

1

2

. The first software unlock triggered the initial price jump a few months before this hardware-level work

2

. Analyst Ken Ashe's verdict was that one unlocked card is a hack while ten repeatable ones would be infrastructure. He called the work "silicon archaeology with a soldering iron"

1

.

© 2026 TheOutpost.AI All rights reserved