Nvidia CMP 170HX: GA100 mining card for USD 5,000

In Nvidia's flagship CMP 170HX the most noticeable were the cooler, the power connection, and the heat spreader that covers the GPU and HBM memory.

Nvidia CMP 170HX: GA100 mining card for USD 5,000

Photo Credits: Linus Tech Tips (Youtube)

Up to now, commercially available graphics cards have mostly been used to mine for cryptocurrencies, but Nvidia now also sells various graphics accelerators specifically for this purpose. The current flagship is the Nvidia CMP 170HX, which is sold for 5,000 US dollars and is based on the 826 mm² GA100.

Lots of unusual solutions https://www.youtube.com/watch?v=EcGkF9SBuSo

There is hardly any official information about the CMP 170HX, but the Youtube channel Linus Tech Tips recently obtained a corresponding model, tried it out, and analyzed it. A trimmed version of the GA100 is used, in which only 4,480 of the maximum 8,192 shaders were activated. In addition, the memory expansion and the PCI-E connection are significantly reduced, but this should be insignificant for the mining performance. Accordingly, there is only 8 GB of HBM memory and a 1.1 interface for the CMP 170 HX.

According to the results from Linus Tech Tips, the CMP 170HX achieves 164 MH / s at 250 W TDP, and with a power limit reduced to 200 W, it is still 160 MH / s. This makes the model much more efficient than, for example, an RTX 3090 (805 kH / watt instead of 264 kH / watt), which is probably due to the high memory bandwidth and manufacturing. The GA100 of the CMP 170HX comes from TSMC's 7 nm process, whereas the GA102 of the RTX 3090 is manufactured by Samsung in their 8 nm process.

The implementation is also exciting about Nvidia's mining graphics card. The model is actually based on a data center GPU, but the implementation is unusual even for such a GPU. For example, there is no normal GPU power plug on the CMP 170HX, but a CPU power plug instead. This is also not soldered directly to the PCB, but rather connected via a cable. The massive heat spreader that covers the large GPU and the surrounding HBM memory is also striking.