CMP 170HX, NVIDIA’s own mining card
Most CMP (CryptoMining Processor) cards were based on board partners’ design, but CMP 170HX is an exception. New photos featuring the design have been shared on Zhihu by Codefordl.
NVIDIA CMP 170HX, Source: Codefordl
Some of our readers might remember that we were first to report on 170HX, an alleged GA100 based CMP model that has not been released alongside Turing/Gaming Ampere mining cards. The 170HX has only just started appearing out in the wild.
The CMP 170HX is a repurposed NVIDIA A100 accelerator, which has no display connectors and no gaming functionality. NVIDIA couldn’t figure out what to do with broken GA100 GPUs, which is why they are now offering this highly advanced Ampere GPU out to miners. It is the first NVIDIA’s own design for a mining card as far as we know.
NVIDIA CMP 170HX has 4480 CUDA cores, which is only 54% of all cores present in the GA100 silicon. For comparison, the GA100 in A100 features 6912 CUDA cores which are 84% of active cores. In terms of memory size, the CMP 170HX has a massive reduction as well. Instead of 40GB or 80GB present in various A100 models, the 170HX only offers 8GB HBM2e memory, supposedly in four stacks (judging from the 4096-bit bus), and a maximum bandwidth of 1.493 GB/s.
NVIDIA CMP 170HX, Source: Codefordl
NVIDIA wanted to ensure that CMP 170HX has no purpose to gamers and researchers, who might have wanted to use its computing power instead of the Data Center Ampere series, which undoubtedly are much more expensive. In order to achieve that, NVIDIA has limited the PCIe interface bandwidth to version 1 and only four lanes (despite physically being 16 lanes). This would be a severe bottleneck for many workloads, just not for cryptomining.
According to GPU-Z data, the GPU has a base clock of 1140 MHz and boost clock of 1410 MHz. The board is detected as Device ID of 20C2 and it is using NVIDIA 471.41 drivers.
NVIDIA CMP 170HX Specifications, Source: Codefordl, @9550pro
The current BIOS version does not support memory overclocking (as of July 28th), which blocks further power optimizations for mining. This is important as memory speed is by far the most important thing in mining.
NVIDIA CMP 170HX Passive Design, Source: Codefordl
In the actual mining test using the Ethash algorithm, the GPU offers up to 164 MH/s without optimizations. Just as much as NVIDIA promised in their official specification datasheet. It is nowhere near 200 MH/s and it might be hard to reach such a hash rate, the author notes.
NVIDIA CMP 170HX Mining Performance, Source: Codefordl
| NVIDIA Crypto Mining Processors | |||||
|---|---|---|---|---|---|
| CMP 30HX | CMP 40HX | CMP 50HX | CMP 90HX | CMP 170HX | |
| GPU | TU116-100 | TU106-100 | TU102-100 | GA102-100 | GA100-100 |
| Board | PG161 SKU 90 | PG161 SKU 100 | PG150 SKU 100 | PG132 SKU 100 | P1001 (?) |
| Memory | 6GB GDDR6 | 8GB GDDR6 | 10GB GDDR6 | 10GB GDDR6 | 8GB HBM2 |
| Rated Power | 125W | 185W | 250W | 320W | 250W |
| Availability | March | March | Q2 | May | Since Q2 |
| Ethereum Hash Rate | 26 MH/s | 36 MH/s | 45 MH/s | 86 MH/s | 164 MH/s |
Source: Codefordl (Zhihu) via @9550pro
