With a little hardware hack to the Nvidia CMP 170HX GPU it’s possible to unlock the full 80GB of VRAM, turning the miner's GPU into a potential option for local AI tasks. This is a huge deal, as it means that we now have access to large amounts of VRAM on the cheap, compared to enterprise-grade options. It will be possible to run bigger models on the cheap.
Hardware mod turns mining GPU into AI powerhouse with 80GB memory
Unlocking the GPU is a hardware hack that requires some modding to break past the factory constraints. This is thanks to a developer (Kate) who managed to pull off the stunt.
Specifications
- VRAM: 80GB HBM2e (unlocked from 8GB)
- SM Count: 120-126 (unlocked from 70)
- FP16 Performance: 85.9 TFLOPS
- Architecture: GA100
CMP 170HX from Nvidia is based on GA100 which is also the silicon behind the A100 pro data center GPU. It was launched in 2021 as an Ethereum dedicated GPU miner accelerator. This GPU was limited to 8GB of VRAM and certain compute units were disabled to distinguish it from more premium offerings. The hardware hack focuses on modding the eFuse controller on the GA100 (Nvidia GPU) using an STM32 microcontroller to glitch security fuses and enable disabled compute resources.
This modification manages to increase the number of Streaming Multiprocessor (SM) from 70 units set by factory to 120-126 units. That's an 80% increase in compute units compared to out-of-the-box units. Also, the new unlocked GPU's reach 85.9 TFLOPS in FP16 testing.
That is about 18% higher than A100 standard, which is 72.5 TFLOPS. The mod requires some soldering and can take from 1h to 48h depending on the GPU binning.
Kate tried to unlock 10 CMP 170HX, they managed to unlock the full 80GB VRAM on 9 out of 10 units. There was one GPU that only managed to reach 64GB, which could suggest that there are some variations or issues when manufacturing. We don't know why Nvidia had heavily disabled this GPU, it's impossible to say if it was manufacturing issues or Nvidia wanted to product segregate them. Thanks to this hardware hack, we can repurpose old mining hardware into AI tasks that need large memory.
It's safe to say that the unlocked version of the CMP 170HX gives a substantial performance upgrade in some use cases. Specifically, tasks that require large VRAM buffers. Local AI inference or training workloads can benefit from this GPU upgrade and now you can have enterprise memory levels on a cheap GPU, thanks to the unlock.
It's a difficult mod that comes with risks, but shows the power of turning specialized hardware into more general computing work. We've been keeping an eye on the CMP 170HX, read our earlier post about NVIDIA CMP 170HX Cracked.



Did you run the unlocked cards through any standard AI benchmarks or just raw TFLOPS counting? Raw compute does not equal usable inference speed.
Finally some cheap VRAM for local models without buying enterprise scrap. My desk is small enough for this.
@Brundle Kate mentioned the FP16 throughput is the primary metric here. The SM unlock from 70 to 126 units drives that 85.9 TFLOPS number. It is a hardware level bypass not a driver optimization.