I can now say with some confidence that the best currently available GPU for mining Grin is the NVIDIA RTX 2080 Ti, with one caveat.
The main reason to prefer the 2080 Ti over the 2080 (and 1080) is the 11 GB memory which allows it to efficiently solve the ASIC targeted PoW, cuckAToo31+.
The reason to prefer the 2080 Ti over the 1080 Ti is the much higher efficiency at running cuckatoo31. Even without using its claimed 64 KB shared memory capability, it’s over twice as fast, and I don’t think its power use is close to twice as high.
The caveat I mention is that when ASICs show up, they may have enough efficiency advantage over the 2080 Ti to relegate it over to the GPU focussed, ASIC resistant PoW, cuckARoo29, where it has to compete with many other cards that only need 6 GB.
In the rare event that ASICs don’t show up in the first year, and cuckatoo31 gets phased out, then the RTX Titan 24 GB looks well positioned to solve cuckatoo32+ efficiently. But at $2499 it’s over twice as expensive as a 2080 Ti…
The reason to prefer the 2080 Ti over the 1080 Ti is the much higher efficiency at running cuckatoo31. Even without using its claimed 64 KB shared memory capability, it’s over twice as fast, and I don’t think its power use is close to twice as high.
Would the same happen with cuckaroo29? Or there’s not really much of a difference at the moment?
Thanks @tromp for the info on this. If this GPU, which costs more than 1000 usd each, is best suited for the Grin POW algo, then how can it be prevented that there is defacto centralization of the mining capacity in Grin?
I am experimenting setting up a Grin miner with commodity hardware, but I guess I will be blown away by these specialized rigs with high-end GPUs, no?
what’s the performance for 1060 card when you run C31 ? just want to check the cost performance and the bottleneck for C31. and is it linear correction or exponential correction with the GPU memory? for example, if i have some Tesla V100 card with 32GB memory, which is like 2.9 times bigger than 2080Ti, when you run C31, the gps should be (1.6 or1.7)2.9 about 4.8 gps or should be like 1.6 (square of 2.9)= 13.4 gps on C31 algorithm? thanks!
A 1060 doesn’t have the 11 GB required to solve Cuckatoo31. A Tesla V100 doesn’t benefit from its extra memory, and is expected to score in between a 1080Ti and 2080Ti. It will however come into its own next year when Cuckatoo31 gets phased out and it will be one of the few cards with the 22 GB needed for Cuckatoo32.