TECH NEWS – While it is not a widely known card, it can still deliver solid results for certain purposes today.

The CMP 170HX GPU was released in 2021. Released at the height of the mining craze, it offered a simplified A100 Ampere configuration with 8 GB and 10 GB VRAM variants. Following the collapse of the cryptocurrency mining market, the card found several alternative uses, but its lack of 3D rendering support meant it was not particularly useful. Five years later, however, it appears that the CMP 170HX has found new life. A new unlocking tool has reportedly been released that removes the limitations of the CMP 170HX chip. It is called CMPUnlocker and is available to download from GitHub.

According to the description, CMPUnlocker restores full SM compute throughput and unlocks the HBM2e memory geometry, which is restricted by Nvidia’s firmware/OTP. Nvidia placed numerous restrictions on the CMP 170HX, effectively rendering it a “card intended solely for mining.” However, the AI era is breathing new life into this chip. There are a few caveats: the 8 GB 170HX can be unlocked to 64 GB and the 10 GB version to 80 GB. Nevertheless, not everyone will see the fully unlocked memory stack, as the CMP 170HX is based on a lower-quality version of the original A100 Ampere GPU, and most of these HBM2e stacks are disabled due to defective or lower-quality memory modules. Consequently, only certain chips will be able to use the full memory capacity, and even then, full memory speed is not guaranteed.

Versions with 32 or 40 GB of memory will also be sufficient, but the maximum capacity can only be achieved with higher-quality memory modules. Some cards are unable to maintain full bandwidth even after the limitation is lifted, reducing clock speeds depending on the quality of the HBM silicon. The bandwidth is approximately 700–800 GB/s per card. The Samsung chips used in the 10 GB models can reportedly be unlocked to 80 GB but are unstable. In contrast, the 8 GB model containing Hynix chips is reportedly more stable when unlocked to 64 GB. Many users are already using the device and running various models on it, so these chips are clearly suitable for certain AI purposes. The chip only supports INT8 data formats; unlike newer Nvidia chips, it does not support FP8 or FP6/FP4, and its performance is estimated at 48 TOPS.

cheap 64GB vram nvidia card?
byu/CabinetNational3461 inLocalLLaMA

According to the original research study, stable operation was only achieved up to 40 GB under load, so reliable operation at 64 GB capacity is not guaranteed. While some users report that the 64 GB version works, others report instability. The device is limited to PCIe Gen2 x4; physical soldering is required to achieve x16 Gen1 and unlock Gen2. As there is no FP8/FP4 support, it is not as fast as a 5090 despite the VRAM. Some cards may have defective memory modules – Nvidia likely disabled the defective parts during testing, so results may vary. The most interesting aspect is the price. Just a few weeks ago, Nvidia CMP 170HX GPUs were available for $100–$200, but following the release of this unlocking tool, they have jumped to over $1,000. Currently, the card is selling for between $1,200 and $2,000 on various eBay listings.

While a 64 GB or 80 GB chip in this price range seems like a great deal on the face of it, you also have to consider the aforementioned drawbacks, which make it something of a lottery. The quality and grade of the chip you receive are never guaranteed, but these chips are currently selling out fast, with some buyers snapping up hundreds at a time.

Source: WCCFTech, GitHub