Nvidia: GA100 GPU used to create ‘a monster for mining’?

Over the past few weeks Nvidia unveiled the range of cards CMP (Crypto Mining Processor), designed specifically for the miner of cryptocurrencies. It is a series of proposals created with the intent of channeling the purchases of the miners in order to preserve GeForce gaming GPUs for their natural audience, that of gamers.

Simultaneously with the creation of this product line, the US company introduced – starting from the GeForce RTX 3060 12GB – a performance limiter to reduce the mining capacity of its gaming GPUs, but only with specific algorithms, such as the one underlying Ethereum. . Unfortunately, as we know, Nvidia has been the protagonist of a slip, with the accidental publication of a driver that allows owners of the new card to undermine at full power.

Nonetheless, Nvidia continues with its plan, and by the end of the second quarter four CMP models will arrive on the market: 30HX, 40HX, 50HX and 90HX. The products are based on a mix of Turing and Ampere GPUs, or rather only the 90HX model based on an Ampere GPU, a variant of the GA102 that we find on board the RTX 3080 and 3090.

There are rumors that Nvidia is going for it working on a fifth model of CMP board, based on GA100 GPU, currently used only in the HPC / IA sector on board the Nvidia A100. To spread the indiscretion il leaker kopite7kimi, which has a past full of “forecasts” that are right.

The very nature of the GPU, designed purely for calculation, could give life to a CMP card that the leaker defines “mining monster“, that is, capable of achieving unprecedented hash rates. There is currently no further information, but we know that the GeForce RTX 3090 reaches an average of 120 MH / s undermining Ethereum via the DaggerHashimoto algorithm.

In addition to the “structure” of the GPU itself (number of cores, etc.), the memory bandwidth, fundamental in the field of mining, and therefore probably a CMP model defined as a “mining monster” should have a particular bus and memory configuration.

The RTX 3090 has 24 GB of GDDR6X memory at 19.5 Gbps on the 384-bit bus for a bandwidth of 936 GB / s, while the CMP 90HX has 10 GB of memory and reaches a hash rate of 86 MH / s. Nvidia hasn’t disclosed all the details of this model, but it is assumed to have a 320-bit bus like the RTX 3080 and use GDDR6 memory instead of GDDR6X.

For its part, Nvidia A100 accompanied by 40 GB of HBM2 memory, capable of offering a bandwidth of 1.6 TB / s on 5120-bit bus. A hypothetical CMP HX card based on the GA100 GPU might not have all that memory, it would be completely useless, but it could keep the bandwidth to really turn into the “mining monster” outlined by the leaker.

In all this discussion, hypothetical for now, one cannot help but also talk about the possible price. An accelerator of that type would be extremely expensive, within the reach of only a certain level of mining farm: the PCI Express version of Nvidia A100, just to put things in perspective, costs around 11 thousand dollars. The alleged CMP incarnation should necessarily cost less, even if only for a lower memory allocation, but the actual value (ie the daily gain in relation to the outlay and energy consumption) would only be understood by the “mining test”, even in relation to dedicated ASICs.