NVIDIA CMP 170HX: From Obscurity to an AI Industry Star

There are truly remarkable transformations in the computer hardware market when products that seemed destined for oblivion suddenly find new life thanks to unexpected innovations. The clearest example of this “second wind” is the NVIDIA CMP 170HX graphics card. Released in 2021 at the height of the mining boom, it was built specifically for cryptocurrency mining, featuring a reduced A100 “Ampere” processor configuration and a limited amount of video memory. However, when the mining hype faded, the CMP 170HX effectively lost its value, especially because it lacked support for 3D rendering, which made it unsuitable for traditional gaming or professional graphics work.

The situation changed dramatically with the appearance of an unofficial software tool called “CMPUnlocker,” which makes it possible to unlock the hidden potential of these graphics cards. This utility, available on GitHub, not only restores the computing power of the streaming multiprocessors (SMs), but also, most importantly, unlocks the HBM2e memory geometry that was artificially limited by NVIDIA firmware. These restrictions were imposed by the manufacturer to turn the CMP 170HX into a mining-only tool. However, the rapid growth and huge demand for artificial intelligence (AI) solutions have opened new horizons for these graphics cards.

Unlocking Hidden Capabilities: Expanded Memory and Its Consequences

CMPUnlocker allows the amount of VRAM to be significantly increased. CMP 170HX versions with 8 GB can be expanded to 64 GB, while 10 GB models can reach an impressive 80 GB. This capability has attracted the attention of professionals and enthusiasts looking for powerful solutions for training large language models, data processing, and other memory-intensive tasks.

However, it is worth noting that memory unlocking does not always go smoothly. The CMP 170HX is based on the A100 Ampere processor, which itself has a lower-level configuration. A significant portion of the HBM2e memory stacks may have been disabled by the manufacturer due to the use of defective or lower-quality memory dies. This means that not every graphics card will be able to fully use the entire unlocked memory capacity. Even in cases where unlocking is successful, full memory speed is not always guaranteed.

Capacities of 32 GB and 40 GB are more achievable, but the maximum 64 GB or 80 GB will only be available on chips equipped with higher-quality memory dies. Some graphics cards after unlocking may not handle the full bandwidth, resulting in clock speed reductions depending on the quality of the HBM chip.

According to reports, Samsung chips used in 10 GB CMP 170HX models have the potential to be unlocked to 80 GB, but they tend to be unstable. In contrast, 8 GB models with Hynix chips prove to be more stable when unlocked to 64 GB. Memory bandwidth per board is approximately 700–800 GB/s.

Despite certain limitations, many users are already actively using the CMP 170HX for AI-related tasks. These graphics cards support the INT8 data format, which is important for many AI models. The performance of the CMP 170HX in such tasks is estimated at around 48 TOPS. However, it is important to remember that stable operation at maximum memory sizes (64 GB) is not guaranteed. Some experimental data show that only about 40 GB is effectively used under load, although some users report stable operation with 64 GB as well.

In addition, the PCIe interface in the original configuration is limited to Gen2 x4 support. Achieving x16 Gen1 and unlocking Gen2 requires physical intervention, namely soldering modifications. It is also important to note the lack of FP8/FP4 support, which makes this card less powerful compared with newer NVIDIA models such as the RTX 5090, despite its much larger VRAM capacity. NVIDIA likely disabled certain memory blocks that did not pass factory testing in order to optimize production costs and avoid defects.

Frenzied Hype and a Price Boom

The most striking aspect of this story is the transformation in the market value of the CMP 170HX. Just a few weeks ago, these graphics cards could be bought for a symbolic $100–$200. However, with the arrival of CMPUnlocker, prices have skyrocketed, surpassing the $1000 mark. On popular online auction sites such as eBay, these models are now listed at prices ranging from $1200 to $2000.

At this price, graphics cards with unlocked memory capacities of up to 64 or 80 GB may seem like an attractive investment for those who need large amounts of VRAM for AI tasks. However, the technical nuances listed above and the potential memory-quality issues turn such a purchase into a real lottery. No one can guarantee exactly which configuration you will receive. Even so, CMP 170HX cards are selling out at lightning speed. Reports are emerging that some buyers are ordering hundreds of units at once, indicating strong demand and a perception of these cards as a strategic resource for AI development. This phenomenon demonstrates how dynamic and unpredictable the technology market can be, where innovations, even unofficial ones, can completely change the value and purpose of products.