NVIDIA's CMP 170HX GPUs have allegedly received a second life with a new tool that unlocks up to 80 GB memory, making them a viable AI solution.
The CMP 170HX GPU was launched all the way back in 2021. It was released at the peak of the mining craze and offered a cut-down A100 "Ampere" configuration with 8 GB and 10 GB VRAM options. The card found multiple uses after the crypto mining crash, but didn't really seem to be of much use , especially due to its lacking 3D rendering support.
Now, five years later, it looks like the CMP 170HX GPUs have found a new life. A new unlock tool has popped up that reportedly unlocks the CMP 170HX chip. This tool is called " CMPUnlocker " and is available on GitHub for download.
| Feature | Status |
|---|---|
| Full SM compute throughput (SS0/SS1) | Working ✓ |
| Memory geometry (64GB on 8GB cards, 40GB on 10GB cards) | Working ✓ |
| PCIe Gen 2 speeds | Working ✓ |
| JTAG (Host2Jtag register access) | Working ✓ |
| Persistence across reboot (patched modules) | Working ✓ |
According to the description, the CMPUnlocker Tool restores the full SM compute throughput and unlocked HBM2e memory geometry that is restricted within the NVIDIA firmware/OTP. NVIDIA placed various locks on the CMP 170HX, making it a "mining only" card, but the AI era is offering a new life to this chip.
There are a few catches, though: the 8 GB 170HX can be unlocked to 64 GB, while the 10 GB can be unlocked to 80 GB but not everyone is going to see the fully unlocked memory stack because CMP 170HX is based on a lower bin of the original A100 Ampere GPU, and most of these HBM2e stacks are disabled due to the use of defective or subpar memory dies. As a result, only certain chips will expose the full memory capacity, and then you aren't assured the full memory speed.
Getting something like 32 GB and 40 GB will be decent enough, but the max memory will only work on chips that have better memory dies. Some cards won't even sustain full bandwidth with the unlock and will lower clock speeds depending on the quality of the HBM silicon. Bandwidth is around 700-800 GB/s per board. It is reported that the Samsung dies used on the 10 GB models unlock to 80 GB but are unstable, while the 8 GB model with Hynix is more stable when unlocked to 64 GB.
Now, whether these chips are practical for AI use, well, many users are already using the tool and running various models. The chip only supports up to INT8 data formats, no FP8 or FP6/FP4 like the newer NVIDIA chips, and the performance is rated at 48 TOPS.
So summing things up (via r/ LocalLLaMA ):
- Stability at 64GB is not guaranteed — the original research paper only got 40GB to work reliably under stress. Some users report 64GB working, others report instability.
- PCIe is limited to Gen2 x4 (you need physical soldering to get x16 Gen1 and unlock Gen2).
- No FP8/FP4 support** — so it's not as fast as a 5090 despite the VRAM.
- Some cards may have defective memory stacks — NVIDIA likely disabled portions that failed testing, so your mileage varies.
The more interesting part is the price. Just a few weeks ago, the NVIDIA CMP 170HX GPUs were selling for around $100-$200 US, and now, the prices have jumped beyond $1000 US as soon as this unlock tool was released. Now, the card is going off for $1200-$2000 in various eBay listings .
Now, if you look at it from one side, a 64 GB or 80 GB chip in this price range is a solid deal, but you also have to consider the drawbacks stated above, which turns this into a lottery. You'll never know what sort of quality or bin you get, but these chips are selling fast now, with some buyers booking multiple 100 units.
Follow Wccftech on Google to get more of our news coverage in your feeds.