I have the 3975wx also … I was using the chia oem app with some scripting in powershell ise. I have around 10 ssd’s 980 pro’s. I am exploring madmax from last night and it seems this is the key i was missing. Would you be able to share any tips as far as thread count per job and stagger times?
1 Like
So I manage to get consistently 65+ plots a day now with quadchannel 128GB RAM, 24C 3960x with 4x2TB NVMe.
The trick seems to be, dont do the AMD hardware/motherboard RAID0, their controllers seem to be a bottleneck for high parallel threading, do individual drives or software striping.
Then its only capped by CPU/IO (so WRX80 next please).
Tune concurrency somewhere above physical core count, 28-36+, go single threaded all the way and let phase 1 share cores.
Thats about 20 to 22 Minutes per plot with about 8-12h ramp up time, which also is the stagger.
With 34 concurrency I get about 6;30, 2:00, 4:15, 0:15. 13 hours.
Still not sure about absolute best concurrency, since that takes a whole 2 cycles/24h to get pure results. Under 28 is certainly leaving CPU underutilized for phase 1.
So 20min single plot time is the time to beat for mad max or other 3rd parties plotters for my setup. Still convinced, that a GPU/CUDA RDMA plotter will eventually make CPU plotting obsolete and might also eat a few k levels, so I’m not keen to invest into a render-beast Threadripper pro I wouldnt know how to use utilize fully yet.
1 Like