With the CMP 170HX, Nvidia is preparing a new mining accelerator based on GA100 with HBM2. Unlike the previous mining graphics cards of the CMP series, the CMP 170HX is not based on conventional Turing or Ampere models from the board partners, but is only manufactured by Nvidia itself – as a Founders Edition, so to speak.
Mining accelerator for server use
The first photos of the CMP 170HX are on the Internet popped upaimed at operators of professional mining farms. The pictures show an approximately 2 slot wide accelerator that dispenses with any fan and is instead cooled passively. The rather long hardware is completely covered by a heat sink, which in turn relies on a good draft, as is usually the case with graphics cards for servers in slide-in format for server racks – the system is reminiscent of the cooler of the also passively cooled Quadro RTX A40 for servers.
HBM2e instead of GDDR6 (X) for mining
Unlike the other models in the CMP series, Nvidia uses the professional GA100 GPU from the Ampere series for the CMP 170HX, which differs from “gaming amps”. The 70 active streaming multiprocessors, GA100 actually carries 128 SMs, are not equipped with 128 FP32-ALUs each, like with gaming amps, but like Turing with their 64 units, so that 4,480 FP32-ALUs are available. For use in the Ethereum blockchain, however, another deviation is much more important: Instead of GDDR6 or GDDR6X, the CMP 170HX uses 8 GB HBM2e.
Nvidia’s professional mining accelerator CMP 170HX (Image: Codefordl)
Image 1 of 6
This clocks at 1,458 MHz, which in connection with the 4,096-bit interface (GA100: 6,144 bits, two HBM2 stacks are missing accordingly) results in a memory bandwidth of 1.5 TB / s. The GPU comes with a base clock of 1,140 MHz and an average boost of 1,410 MHz. So that the mining accelerator is not used as a slower, but much cheaper professional graphics card, Nvidia has apparently limited the PCIe interface to 4 PCIe lanes of the first generation. This incision and the limitation in the clock rate are not a problem for the intended purpose, because the memory bandwidth is what counts in the first place.
The TDP of the CMP 170HX is then only given as 250 watts. Apart from the PCIe slot, these are provided via an 8-pin CPU connector, which is connected to two 8-pin PCIe connectors by means of an adapter.
The desired result is achieved
In an initial test, the Nvidia CMP 170HX achieved a result of 164 MH / s in the Ethash algorithm, so Nvidia’s naming, which promises 170 MH / s, is appropriate. Neither the GPU nor the memory can currently be overclocked. If the power limit is reduced to 190 watts, the hash rate should hardly drop at 158 MH / s, so efficiency would continue to increase.
GPU-Z screenshot from Nvidia CMP 170HX (Image: Codefordl)
Post a Comment
Post a Comment