Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

Hacker Kate Turns a Mining CMP 170HX Into an 80 GB A100, 20% Faster Than the Retail Card

A mining NVIDIA CMP 170HX with 8 GB of memory has been turned into a full-fledged 80 GB A100. Researcher Kate dropped the GA100's supply voltage at the moment of power-up, when the chip reads the fuses that hold the factory restrictions, and it wrongly concluded that all of them were intact. The full 80 GB of memory, PCIe Gen4, NVLink and MIG now work, and the SM count rose to 120–126 versus 108 on the retail A100. The takeaway: NVIDIA's hardware restrictions can be bypassed, though not everyone will be able to repeat it.

AI-processed from TG: @derplearning (Derp Learning); edited by Hamidun News
Hacker Kate Turns a Mining CMP 170HX Into an 80 GB A100, 20% Faster Than the Retail Card
Source: TG: @derplearning (Derp Learning). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @derplearning (Derp Learning) ↗
Editor-in-Chief
CMP 170HX8 GB of memory, NVLink disabled, PCIe limited to Gen1
Key metric #2120–126 working SMs versus 108 on the retail A100: roughly +20% performance
Separatelyfirmware flashing on GeForce 3xxx/4xxx cards yielded almost 2x TFLOPS

When NVIDIA sells identical silicon at different price points, the difference between the products is often not written into the chip itself, but into settings the chip reads at power-up. Researcher Kate showed just how fragile this kind of protection can be: she took a cut-down 8 GB CMP 170HX mining card and turned it into a full-fledged 80 GB A100. Its performance ended up higher than that of the retail version.

How NVIDIA Turned a Top-End Chip Into a "Miner"

During the crypto boom, miners were buying up graphics cards, and NVIDIA decided not to hand them data-center hardware. It took a healthy GA100 chip, the one the A100 is built on, and used it to make the CMP 170HX card. The restrictions were set at the level of the chip itself, in electronic fuses (eFuses). These are tiny links that are "blown" at the factory and cannot be restored afterward.

What exactly was cut:

  • memory was physically cut down to 8 GB;
  • NVLink, the high-speed bus that connects cards to each other, was disabled;
  • PCIe was reduced to Gen1;
  • tensor compute throughput was blocked, meaning the matrix math that AI accelerators are bought for in the first place.

That was enough for mining, but it made the card barely usable for training and running models.

How the Hack Works: Fooling the Chip in the First Milliseconds

The idea resembles a document check at an entrance. A guard looks at a list, you cut his light for a fraction of a second, and he writes "all in order" in the log. Kate found that the GA100 is vulnerable to Voltage Fault Injection, that is, fault injection through the power supply. At startup, the chip reads the fuse configuration and verifies it. The window in which this happens is very short. At that moment Kate fed the power line a precisely calculated pulse (voltage glitching) and disturbed the fuse sense amplifiers. The amplifiers started making errors, and during the self-check the chip took all the blown fuses for intact ones.

Hardware work was needed as well: high-speed memory was attached to the chip, with its solder balls restored from scratch (reballing). Together, these steps removed all the factory restrictions. The memory was detected as 80 GB, and the previously crippled PCIe Gen4, NVLink and MIG, the technology for slicing one card into several isolated ones, started working at full capacity.

Bonus: The Card Turned Out Faster Than the Retail A100

The most unexpected result came next. To guarantee a good yield of working dies, NVIDIA enables only 108 streaming multiprocessors (SMs) on retail A100s, even though the die has more. This is normal practice in chip manufacturing: defective blocks are disabled and spare ones are kept in reserve, so as many dies as possible can be sold. After the hack, all the functional blocks on Kate's chip came alive, from 120 to 126 of them. According to the author, the real compute performance came out roughly 20% higher than that of the official A100.

The Derp Learning post also mentions a related case. Enthusiasts used firmware to unlock GeForce 3xxx and 4xxx chips and got almost a 2x increase in TFLOPS. But memory speed stayed the same, so the speedup is not 2x for every operation: where a workload is memory-bound, the gain is much more modest.

What This Means in Practice

First, don't count on cheap A100s from decommissioned mining rigs. A hack like this requires an oscilloscope, precision equipment for delivering the pulses, BGA memory soldering skills, and a willingness to kill the card if an attempt fails. This is a lab experiment, not a way to refresh a GPU fleet. The retelling contains no data on long-term stability or on how such a card behaves under real workloads.

Second, for the security industry this is one more illustration that protection written into silicon is not permanent. If the entire restriction logic relies on the chip honestly reading its own fuses at startup, it is enough to disrupt its power at the right moment. Manufacturers will have to think about resistance to power-supply faults, and hardware enthusiasts can expect more stories about unlocking consumer and data-center cards.

Third, for those building infrastructure, the conclusion is simple: chip-level restrictions are put in place by the manufacturer for business-model reasons, not because the hardware is incapable of more. Understanding where the line runs between what the die can actually do and the limits set by marketing helps you judge accelerator prices soberly. Details and technical specifics are available on Kate's blog, linked in the original source.

Story

Key Questions Answered

How NVIDIA Turned a Top-End Chip Into a "Miner"?

During the crypto boom, miners were buying up graphics cards, and NVIDIA decided not to hand them data-center hardware. It took a healthy GA100 chip, the one the A100 is built on, and used it to make the CMP 170HX card. The restrictions were set at the level of the chip itself, in electronic fuses (eFuses). These are tiny links that are "blown" at the factory and cannot be restored afterward.

How the Hack Works: Fooling the Chip in the First Milliseconds?

The idea resembles a document check at an entrance. A guard looks at a list, you cut his light for a fraction of a second, and he writes "all in order" in the log. Kate found that the GA100 is vulnerable to Voltage Fault Injection, that is, fault injection through the power supply. At startup, the chip reads the fuse configuration and verifies it. The window in which this happens is very short.

What are the key details on Bonus: The Card Turned Out Faster Than the Retail A100?

The most unexpected result came next. To guarantee a good yield of working dies, NVIDIA enables only 108 streaming multiprocessors (SMs) on retail A100s, even though the die has more. This is normal practice in chip manufacturing: defective blocks are disabled and spare ones are kept in reserve, so as many dies as possible can be sold. After the hack, all the functional blocks on Kate's chip came alive, from 120 to 126 of them.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

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.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…