Page 2 of 2
Re: [Windows] RX 590: Particle coordinate is nan (Resolved)
Posted: Sat May 01, 2021 5:33 pm
by gunnarre
Projects with fewer atoms being simulated might not utilize all the compute units in a particular GPU. There is work under way to match the project to the GPU more closely. While not fully there, the Covid Moonshot is split into two tiers for small and large GPUs.
Re: [Windows] RX 590: Particle coordinate is nan (Resolved)
Posted: Sat May 01, 2021 6:20 pm
by bruce
Talgeeze wrote:The only odd thing I noticed is, for some WUs, the average GPU usage is quite low (~45%) ....
I expect that it's those proteins with smaller numbers of atoms. That GPU does best on larger proteins. If you give us a list of the projects, we might get the project scientist to configure his project to exclude assigning those proteins to GPUs like yours.
There's a project to optimize the assignments to make better use of the GPUs but it seems to be stalled.
Re: [Windows] RX 590: Particle coordinate is nan (Resolved)
Posted: Sat May 01, 2021 9:19 pm
by Talgeeze
bruce wrote:
If you give us a list of the projects, we might get the project scientist to configure his project to exclude assigning those proteins to GPUs like yours.
Project 13446 would be one.
I'll keep track over the next few days for others like it, currently on 17434 and it has been sitting around 65 % GPU use for the first ~30% of the current WU then jumped to ~80 % and ETA was cut by like 25 %.