That is because I won't be buying anything until the 5 nm generation comes out and the chip shortage is eased. It won't be for a couple of years I think.
Is there a way to speed the toplist calculation up on Windows? It is quite hard to me to fathom why the OS would limit the performance of the Win app so much compared to the Linux app. Why does sth that takes mere sec on a Linux host, take several min on a comparable Win host? The performance impact is huge as the GPU sits idle nearly a third of the computation time of a WU on my Win host.
Any idea of how I could speed this up would be much appreciated!
Yeah with O3 Workunits I use multiple GPU tasks to compensate the post-processing at 90%. After reboot, I just keep them at around 30%+70%+90% so utilization is always ~100%, It's a bit tedious because other than suspending tasks to get the spread, there is nothing else you can do. But I don't reboot that often.
FGRP tasks are much easier to use. One task one GPU, about 1-2sec without utilization once task is completed and until next task starts.
Yeah, that's what I tried so far Solling2 but the GPU sat idle for as long as 5-6 min once the 2 concurrently running WUs both reached the CPU only toplist calculation part. So I did try to even run more than 2 concurrently as the GPU CUDA load for only one O3 GW task is only roughly 70-80% on my 1660S card. Running the setup now with 3 tasks concurrently helps to keep up the load on the card. Thanks for that advice! It does work epecially well when paired with the manual supsend/resume intervention as EXARD3K suggested. Thanks for that suggestion as well! I am however beyond the babysitting of tasks and running occassionally into idle GPU periods albeit short ones, I guess, is ulitmately unavoidable. I just let them run as is once the first 3 tasks are set up to run at different progress levels, so one task should always be running to keep the load on the GPU.
Nevertheless would be interested in understanding why Linux is so much more efficient in the toplist calclulation part of the application. :)
mikey wrote:The really good
)
I just today looked up the AMD GPU roadmap.
https://www.wepc.com/news/updated-amd-roadmap-slides-confirm-5nm-zen-4-cpus-and-rdna-3-gpus-by-2022/
That is because I won't be buying anything until the 5 nm generation comes out and the chip shortage is eased. It won't be for a couple of years I think.
My RX 570 looks just fine until then.
Is there a way to speed the
)
Is there a way to speed the toplist calculation up on Windows? It is quite hard to me to fathom why the OS would limit the performance of the Win app so much compared to the Linux app. Why does sth that takes mere sec on a Linux host, take several min on a comparable Win host? The performance impact is huge as the GPU sits idle nearly a third of the computation time of a WU on my Win host.
Any idea of how I could speed this up would be much appreciated!
bozz4science schrieb: ... as
)
That's what we really want to avoid. Did you try running two (or even three?) tasks at a time, in a staggered manner? :-)
Yeah with O3 Workunits I use
)
Yeah with O3 Workunits I use multiple GPU tasks to compensate the post-processing at 90%. After reboot, I just keep them at around 30%+70%+90% so utilization is always ~100%, It's a bit tedious because other than suspending tasks to get the spread, there is nothing else you can do. But I don't reboot that often.
FGRP tasks are much easier to use. One task one GPU, about 1-2sec without utilization once task is completed and until next task starts.
Yeah, that's what I tried so
)
Yeah, that's what I tried so far Solling2 but the GPU sat idle for as long as 5-6 min once the 2 concurrently running WUs both reached the CPU only toplist calculation part. So I did try to even run more than 2 concurrently as the GPU CUDA load for only one O3 GW task is only roughly 70-80% on my 1660S card. Running the setup now with 3 tasks concurrently helps to keep up the load on the card. Thanks for that advice! It does work epecially well when paired with the manual supsend/resume intervention as EXARD3K suggested. Thanks for that suggestion as well! I am however beyond the babysitting of tasks and running occassionally into idle GPU periods albeit short ones, I guess, is ulitmately unavoidable. I just let them run as is once the first 3 tasks are set up to run at different progress levels, so one task should always be running to keep the load on the GPU.
Nevertheless would be interested in understanding why Linux is so much more efficient in the toplist calclulation part of the application. :)