Rumors of a possible Nvidia CMP 170HX surfaced a few months ago, and although Nvidia hasn’t officially announced it yet, this new mining-oriented card has already been photographed and tested, confirming its existence and specifications.
From these pictures, it appears that the CMP 170HX will not be assembled by AIBs like the rest of the CMP lineup, but by Nvidia itself. It will be based on the reference design of the A100 accelerator and will have the exact same PCB and cooling, though not the same specs, which differ significantly for both the GPU and RAM.
According to a GPU-Z screenshot of this leaked model, this new mining GPU has only 4480 Cuda cores from the 8192 of the GA100 core, which is a drastic reduction compared to the 6912 Cuda cores of the A100 accelerator. This is possible because cores are being used that do not meet the necessary validations for use in the A100 accelerator and are therefore being redirected to this new mining card instead of being discarded.
On the RAM side, it only has 8 GB of RAM memory instead of the 40 GB that the A100 accelerator has, although HBM2 allows for an impressive hash rate of almost 170 MH/s thanks to its 1.5 TB/s bandwidth.
Nvidia’s latest move is to downgrade the interface to PCI-E 1.0 x4 instead of PCI-E 4.0 x16, ensuring that no bandwidth is available for tasks that require the much more expensive professional GPUs.
According to the initial mining tests, this GPU achieves 164MH/s with a power consumption of 250W, practically reaching the 170MH/s it should achieve according to its own name. It is important to note that its bios do not allow memory overclocking. So if new bios are released that support this, its performance could be even higher.
There are no details yet on the availability of the GPU or its price, but we’ll keep you updated as soon as those details are known. At the moment, Nvidia hasn’t even officially announced it yet, but if samples are already in circulation, we could see an announcement soon.