I run the app currently on my notebooks only. One of them has a build-in GTX470m. The GPU usage never exceeded 43% with very long periods of 1 - 2 %.
Checkpointing seems not to be the final solution; I stopped crunching with a wu @ 88,888% with a remaining time of 25 min.
The work distribution does not work according the preferences. Also I've set to get FGRP #3 only, I got other wu's too, most of them casa's. And the GPU utilisation factor of 0.5 is also ignored. With a 'memory used' of only 127MB I could run easily 5 or 6 of these wu's. But this would require 6 or 8 core cpu's or in other words FGRP#3 only pc's.
If this is the final version according to the gpu usage a way must be found to run a stable combination of let's say 2 perseus arm wu's together with one FGRP#3, otherwise the GPU would run idle for longer periods.
Another way would be to make the gpu utilisation factor an argument for the wu type rather than to the pc's venue.
Alexander
Copyright © 2024 Einstein@Home. All rights reserved.
FGRP #3 feedback
)
Result for FGRP #3: 3060s, i5-3570K @ 4.2 GHz, no GPU
BTW: Is AVX support (also for S6-CasA etc.) possible?
For example, one WU Asteroids@home (MacBookPro Late 2013 2.4GHz Haswell):
110 min OSX-SSE3 and
90 min OSX-AVX.
HNY Jürgen.
RE: BTW: Is AVX support
)
MrS posted this idea @Einstein in the Android thread.
I think it's a good idea since more and more cpu's have this feature and a lot of resources are wasted by not using these possibilities.
On the other hand we should be aware that this project is really driven very professional; they work on an improved algorithm, cuda 5.5, have arm wu's for Linux and Android, support for all types of GPU's, have plans to introduce the wisdom file for the fft for different platforms aso. They are working hard to improve the project, which is the core for the success here (Remember: the PetaFlop barrier is crossed and stable above 1 PF)
Alexander
RE: I run the app currently
)
Same here....
My host
)
My host https://albertathome.org/host/9923/tasks (GTX 560 non OC'd with 301.42 driver, i7 860 @2.8Ghz) calculated 29 FGRP#3 FGRPopencl-nvidia tasks with av. of 650 sec of Run/CPU time (A@H was the only project being calculated at that time). All WUs finished without error, but 23 of them validated so far as "Validate error" and 6 are still pending. Can you take a look into these results?
Regards
Luke
Nice work. Please add an
)
Nice work.
Please add an option "Use Intel-GPU" for FGRPopencl-intel_gpu-lion. Albert sent my Host 10045 (OS X) only cpu-workunits.
Jürgen
Now everything runs smoothly
)
Now everything runs smoothly and without "validate error" problems, on both Nvidia GTX560, AMD HD 7970 and different CPUs (Phenom II X4, i5 Ivy and i7 Lynnfield).
I hope we can test production version of FGRP #3 soon (I mean "real" samples with full GPU load).
GTX560
HD7970
Regards
Luke
Well, the FGRP GPU app is in
)
Well, the FGRP GPU app is in the first stage of development, similar to the first BRP GPU App: only the FFT is ran on the GPU, all other computation is still done on the CPU. For the time being we are more concerned to get it running at all and give the same results as the CPU version (i.e. validation). Further improvement is planned with time.
The text for the "GPU utilization factor" preference setting actually reads "GPU utilization factor of BRP apps", which obviously does not apply to FGRP apps. We'll add another such setting for FGRP.
BM
BM
One of my hosts today
)
One of my hosts today received an FGRP3 opencl_nvidia task
It progressed from 0 to 97% indicated completion over about a 50 minute period, with very sporadic indicated GPU usage ranging from 0 to perhaps 60%.
Subsequently it sat at 97.297% indicated completion for over twenty-five wall-clock minutes, steadily using one core of CPU, no discernible GPU, and no indicated progress.
I've suspended it for the time being, and am willing to give it more time if it seems likely to generate some useful information by doing so, but it seems likely it has hung up.
The host is a Windows 7 machine with a Xeon E5620 CPU running hyperthreaded. The graphics processor is a GTX660 with 327.23 driver installed.
That was quite normal when I
)
That was quite normal when I ran both Nvidia and ATI tasks, it'll finish eventually,
Claggy
Claggy wrote:it'll finish
)
I turned it back on, and it finished, after spending well over half as long in the terminal phase of no visible progress and no apparent GPU use as it did for the first 97% of reported progress. Thanks for steering me to continue.