Alright I'm donating compute performance to Leela Chess Zero, a open source project to make a chess playing neural network similar to googles Alpha Zero.
But it's runnly extremely poorly on my RX 480 compared to on Nvidia due to CUDNN being used to accelerate the code on Nvidia hardware.
A 1070 gets about 6500 nodes pr second at stock speeds right now while my 480 is only getting about 400 to 500 most of the time with a few peaks up into 5000 nodes pr second for individual moves...
So I'm wondering is there some form of AMD alternative to CUDNN out there that I could point towards when talking with the devs?
Ideally something that's low level enough to use all of AMDs features.
Because right now anyone who are considering buying new hardware to contributing to that project or for that matter use the neutral network for their own chess games will probably gravitate towards Nvidia for their GPU acceleration...
And considering AMDs GPGPU capabilities I'm sure there's nothing stoping the hardware from being as capable as the competing GPUs.
Any particular page on the GPUOpen where I can get the relevant information or any particular libraries I can point to?
hipDNN seems to be low on developers though, so I hope someone can point more developers in hipDNNs direction so perhaps it'll look attractive to potential neural network devs. =)