The NVIDIA CMP 170HX are probably some of the most peculiar cards you can currently find for building a local artificial intelligence system. They were originally designed for cryptocurrency mining and NVIDIA deliberately limited both their accessible memory capacity and part of their compute performance.
The interesting thing is what they hide underneath: NVIDIA GA100 silicon, the same architecture used by the NVIDIA A100.
After discovering the community unlock project for these cards I decided to buy two different units: an 8 GB CMP 170HX and a 10 GB CMP 170HX.
The result has been much better than I expected.
Two CMP 170HX and 104 GB of VRAM
There are two especially interesting variants of the CMP 170HX:
| Variant | Original VRAM | Unlocked VRAM |
|---|---|---|
| CMP 170HX 8 GB | 8 GB | 64 GB |
| CMP 170HX 10 GB | 10 GB | 40 GB |
| Total of my two cards | 18 GB | 104 GB |
The difference between the two is not simply the amount of memory NVIDIA decided to show the system.
The 8 GB variant uses a different memory configuration and the currently documented stable unlock allows accessing 64 GB. The 10 GB variant reaches 40 stable GB.
80 GB configurations have also been experimented with on the 10 GB version, but they should not currently be considered usable: although the GPU can come to report that amount, errors appear when using memory above approximately 40 GB.
That is why I have preferred to keep my cards in the configurations that are giving me stability: 64 GB + 40 GB = 104 GB of VRAM.
What is a CMP 170HX really?
Here is the most interesting part.
The CMP 170HX uses the GA100, the huge 7 nm chip developed by NVIDIA for the A100 family. In the CMP we find a cut-down version of the same silicon, with 70 SMs, 4,480 FP32 cores and Compute Capability 8.0.
This does not magically turn a CMP 170HX into an A100. There are important differences in configuration, connectivity, number of active SMs and other characteristics.
But for certain compute workloads —and particularly for my goal, which is to run AI models locally— having Ampere architecture, HBM memory and tens of gigabytes of VRAM makes these cards extremely interesting.
The unlock
The community project cmp170hx / cmpunlocker has managed to remove, through software, several of the restrictions imposed on these cards.
The current procedure works over the open modules of the NVIDIA driver and applies the changes during GSP boot. It does not require physically modifying the memory nor converting the card through a simple VBIOS change.
Among other things, the unlock modifies the memory geometry used by the controller and removes the SM compute performance limitation.
In the variants I have:
- CMP 170HX 8 GB → 64 GB
- CMP 170HX 10 GB → 40 GB
The unlock is reapplied when the driver loads, so we are not talking simply about a permanently modified BIOS.
And how do they work for AI?
This was for me the really important question.
My goal is not mining nor using these cards as graphics GPUs. I bought them specifically to experiment with local language model inference.
And in that territory the result is being surprisingly good.
I currently have 104 GB of stable VRAM between the two CMP 170HX, which allows me to run models that normally require considerably more expensive professional hardware.
With llama.cpp I can split the layers and weights between both GPUs, even though they have different amounts of memory. This is especially interesting in my setup because I can take advantage of the 64 GB of one card and the 40 GB of the other asymmetrically.
For large GGUF models, that flexibility is fundamental.
I have been able to experiment with models of tens of billions of parameters and relatively high quantizations without having to offload a significant part of the model to RAM. And when necessary, llama.cpp allows combining VRAM and system memory.
Not everything is advantages
The CMP 170HX are not a perfect solution either.
They are cards designed for mining and lack some features we would expect from a conventional GPU. In addition, their PCIe connectivity is limited and the unlock ecosystem continues to be experimental.
They also need adequate cooling. They are GA100 GPUs and, although for inference they work very well in my experience, you have to pay attention to temperatures, power and airflow.
The unlock also does not turn all physically disabled components into functional ones.
Therefore, I would not recommend buying a CMP 170HX simply thinking you are buying a cheap A100.
It is a peculiar, experimental platform that requires being willing to work with Linux, drivers, kernel modules and tools that are still evolving.
Was it worth it?
In my case, yes.
I bought them as part of my personal laboratory to experiment with local AI and they have ended up becoming one of the most interesting pieces of the equipment.
Going from two cards originally marketed with 8 and 10 GB to having 64 and 40 usable GB respectively completely changes their possibilities for inference.
Having 104 GB of GA100-based VRAM in a personal machine allows experimenting with models that normally fall outside the reach of a conventional PC.
And possibly the most interesting thing is that the project is still evolving.
For someone interested in running and experimenting with local AI models, the CMP 170HX have become much more interesting hardware in 2026 than NVIDIA probably imagined when it launched them as mining-dedicated cards.