I ran comparison tests of 1X productivity on the Windows beta 1.28 application for AMD GPUs on two systems.
In both cases I ran about 5 tasks at 1X on 1.28, then about nine on 1.22, followed by another 5 on 1.28.
For this system running an RX 5700 card I observed about a 3% improvement in productivity for the 1.28 beta code over the 1.22 production code. The task to task variation in elapsed time was quite small, and I am very confident there was a real difference, and even that the magnitude is not greatly different from 3%.
For another system running an RX 6800 card I observed about one tenth as much improvement. Also the task to task variation was more, so I am not confident as to which side of zero the improvement lies, but pretty sure is is very near break even indeed.
I think the question of whether multiplicity even matters is not generally settled, though I suspect it does in some cases. I think the general picture of very modest improvement in some cases, and none in other cases persists, even if somehow one can dispense properly with all reports of difficulty.
I had a similar problem a couple of weeks ago with my Nvidia card. Did you update/change drivers/system? I did and didn't reboot afterwards which caused all kinds of trouble including dumping aborted tasks en masse on the BOINC Client.
"Have you tried turning it off and on again?" Solved it for me.
I only grabbed 10 tasks and turned off beta as soon as I saw they were aborting.
I need more location profiles. Between a 0% backup GPU, 100% GPU Beta, 100% GPU non-beta, 100% BRPS on RPI and 0% GRPS #5 CPU there aren't enough profiles to go around. I guess I'm only have E@H as backup on the NV GPUs.
With ~40 task for each version in 2x parallel mode I see a 1 percent advantage for 1.28 which I did not see in 4x mode.
With 4 parallel tasks I get a 1 percent speed increase with 1.22 and 1.28 over 2x 1.28.
All this is easily diminished by the GPU idle time (~8 sec) that is more often occuring in 2x parallel or 1x mode. So for me it is back to 4 task in parallel with interleaved starting times.
I ran comparison tests of 1X
)
I ran comparison tests of 1X productivity on the Windows beta 1.28 application for AMD GPUs on two systems.
In both cases I ran about 5 tasks at 1X on 1.28, then about nine on 1.22, followed by another 5 on 1.28.
For this system running an RX 5700 card I observed about a 3% improvement in productivity for the 1.28 beta code over the 1.22 production code. The task to task variation in elapsed time was quite small, and I am very confident there was a real difference, and even that the magnitude is not greatly different from 3%.
For another system running an RX 6800 card I observed about one tenth as much improvement. Also the task to task variation was more, so I am not confident as to which side of zero the improvement lies, but pretty sure is is very near break even indeed.
I think the question of whether multiplicity even matters is not generally settled, though I suspect it does in some cases. I think the general picture of very modest improvement in some cases, and none in other cases persists, even if somehow one can dispense properly with all reports of difficulty.
https://einsteinathome.org/ta
)
https://einsteinathome.org/task/1176301809
These just abort on my Radeon VII in Win10
mmonnin
)
I had a similar problem a couple of weeks ago with my Nvidia card. Did you update/change drivers/system? I did and didn't reboot afterwards which caused all kinds of trouble including dumping aborted tasks en masse on the BOINC Client.
"Have you tried turning it off and on again?" Solved it for me.
mmonnin
)
known problem. the Windows drivers don't seem to support this app with Vega or earlier.
turn off beta tasks to solve the issue for Windows/Vega so you do not get this app.
_________________________________________________________________________
I only grabbed 10 tasks and
)
I only grabbed 10 tasks and turned off beta as soon as I saw they were aborting.
I need more location profiles. Between a 0% backup GPU, 100% GPU Beta, 100% GPU non-beta, 100% BRPS on RPI and 0% GRPS #5 CPU there aren't enough profiles to go around. I guess I'm only have E@H as backup on the NV GPUs.
On my 5700X with four
)
On my 5700X with four parallel FGRP gpu task and nothing else I see no difference between 1.22 and 1.28 beta at all. (Running Win 10 21H1 64bit.)
NoxCivi wrote: On my 5700X
)
What about 2x? I don't seem to gain anything at 3x on a couple of Rx 5700's.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
With ~40 task for each
)
With ~40 task for each version in 2x parallel mode I see a 1 percent advantage for 1.28 which I did not see in 4x mode.
With 4 parallel tasks I get a 1 percent speed increase with 1.22 and 1.28 over 2x 1.28.
All this is easily diminished by the GPU idle time (~8 sec) that is more often occuring in 2x parallel or 1x mode. So for me it is back to 4 task in parallel with interleaved starting times.