Currently running at 0.33 (aka x3) and seem to recall even successfully (albeit slower) running at x6 and x8, now with amdgpu-pro-16.50-362463.
Yes, I did a search before posting this thread, and found your old posts where you mentioned running 8 tasks in parallel, which made me all the more curious why 2 tasks were running so slow on my Windows machine.
I hope the drivers will be improved in the future to allow even more efficient use of the RX 480 on Windows.
I hope the drivers will be improved in the future to allow even more efficient use of the RX 480 on Windows.
I hope so too. I'm a little surprised that other RX 480 users running Windows haven't pointed this out before now. You're using Windows 10, as is the host Gavin linked - I wonder if it's just a Win10 thing. Might be worthwhile seeing if there's a different driver available to try.
also... and with a little help I could switch another dual 480 host to Kubuntu and explore the benefits of Linux and AMD GPU PRO drivers in the hope of unleashing these cards true potential!
You know you want to.... penguins are friendly.
I have some RX 460s arriving in the next couple of days. I was going to look at working out how to get them going under PCLinuxOS which doesn't have any amdgpu-pro stuff in its repo. Whilst I'd really prefer to keep everything on PCLinuxOS, it's probably the KDE desktop that makes it so easy, so I might just download Kubuntu and give it a go. Maybe I can configure it to behave exactly like PCLinuxOS :-).
I have six Q6600 quad core hosts (2008 vintage - no GPUs) gathering dust at the moment. They never did get restarted when I fixed up the rest of the fleet from the storm damage of last December. I have about 20 CPU only hosts in total in that 'currently retired' category :-). There's also about a further 10 with GTX650s that will remain off for the time being until we see if some kind soul ports the app to CUDA.
So I agree entirely with AgentB - you just need to do it :-). And yes, penguins are super cool as well as friendly :-).
I'll probably start a new thread about getting RX 4xx GPUs running well under Kubuntu - if I manage to get that far :-).
I have just replaced the Quadro+Tesla K20 with RX 480.
First run with stock settings, 1WU took ~660s.
Then I tried to run 2 WUs and I can see the same problem: GPU Memory Controller utilization very low, GPU Power only ~23W most of the time with small spikes to ~120W at same time as Memory Controller utilization spikes.
It seems that both Polaris and Fiji are unable to run >1WU efficiently, probably a driver problem.
Does anyone know whether we can inform AMD of this problem (and how), or would that be a futile exercise?
I have a good partner contact there, so I can try. But I think the chances of fixing this are rather low as their priorities are probably gaming-oriented and this is a pretty specific problem, probably not worth their effort.
This might also be some application-specific problem, since BRP4G/BRP5/BRP6 didn't exhibit such problem, though there were other issues mostly causing a high number of invalids when running multiple WUs.
Have a RX480 running under W7 and it is around 660 sec single task.
Just replaced it first with a W9100 which is supposedly a monster in DP (double precision) but it ran about the same times as the RX480 also using the 1.18 beta app. I thought that new beta was optimized for DP. Then tested a W8100 and getting the same times again, around 660 sec and all this on the same computer. The Fire Pro's (W9100 and W8100) take at least 40-60 Watts more.
So why am I not getting the DP performance out of these Wx100 cards while the RX480 Polaris chip should only be good for SP?
If you see "OpenCL device has FP64 support" in the logs, it means that the GPU has been recognized to support double floating point. Don't worry about performance, double precision is not the major part of processing.
Yeah, that was my experience too - I don't see any significant improvement on GPUs with good (1/2) FP64 support. Even the v1.18 beta which promised improvement on FP64 didn't meet the expectations. SP code seems to be what matters most, so a Fury X or RX 480 with poor FP64 still perform better than Hawaiis which had good FP64.
I measured 660 secs on RX 480 too, but my CPU was heavy occupied with other CPU tasks. Suspending all other tasks and letting the CPU thread supplying GPU reach max turbo clock can yield ~520 secs there.
So CPU performance still plays a significant role, more significant than good FP64 support.
AgentB wrote:Gavin_14
)
Yes, I did a search before posting this thread, and found your old posts where you mentioned running 8 tasks in parallel, which made me all the more curious why 2 tasks were running so slow on my Windows machine.
I hope the drivers will be improved in the future to allow even more efficient use of the RX 480 on Windows.
Michgelsen wrote:I hope the
)
I hope so too. I'm a little surprised that other RX 480 users running Windows haven't pointed this out before now. You're using Windows 10, as is the host Gavin linked - I wonder if it's just a Win10 thing. Might be worthwhile seeing if there's a different driver available to try.
Cheers,
Gary.
AgentB wrote:Gavin_14
)
I have some RX 460s arriving in the next couple of days. I was going to look at working out how to get them going under PCLinuxOS which doesn't have any amdgpu-pro stuff in its repo. Whilst I'd really prefer to keep everything on PCLinuxOS, it's probably the KDE desktop that makes it so easy, so I might just download Kubuntu and give it a go. Maybe I can configure it to behave exactly like PCLinuxOS :-).
I have six Q6600 quad core hosts (2008 vintage - no GPUs) gathering dust at the moment. They never did get restarted when I fixed up the rest of the fleet from the storm damage of last December. I have about 20 CPU only hosts in total in that 'currently retired' category :-). There's also about a further 10 with GTX650s that will remain off for the time being until we see if some kind soul ports the app to CUDA.
So I agree entirely with AgentB - you just need to do it :-). And yes, penguins are super cool as well as friendly :-).
I'll probably start a new thread about getting RX 4xx GPUs running well under Kubuntu - if I manage to get that far :-).
Cheers,
Gary.
I have just replaced the
)
I have just replaced the Quadro+Tesla K20 with RX 480.
First run with stock settings, 1WU took ~660s.
Then I tried to run 2 WUs and I can see the same problem: GPU Memory Controller utilization very low, GPU Power only ~23W most of the time with small spikes to ~120W at same time as Memory Controller utilization spikes.
It seems that both Polaris and Fiji are unable to run >1WU efficiently, probably a driver problem.
-----
Does anyone know whether we
)
Does anyone know whether we can inform AMD of this problem (and how), or would that be a futile exercise?
Michgelsen wrote:Does anyone
)
I have a good partner contact there, so I can try. But I think the chances of fixing this are rather low as their priorities are probably gaming-oriented and this is a pretty specific problem, probably not worth their effort.
This might also be some application-specific problem, since BRP4G/BRP5/BRP6 didn't exhibit such problem, though there were other issues mostly causing a high number of invalids when running multiple WUs.
-----
https://community.amd.com/com
)
https://community.amd.com/community/devgurus/opencl, however this is a forum for developers mainly.
Have a RX480 running under W7
)
Have a RX480 running under W7 and it is around 660 sec single task.
Just replaced it first with a W9100 which is supposedly a monster in DP (double precision) but it ran about the same times as the RX480 also using the 1.18 beta app. I thought that new beta was optimized for DP. Then tested a W8100 and getting the same times again, around 660 sec and all this on the same computer. The Fire Pro's (W9100 and W8100) take at least 40-60 Watts more.
So why am I not getting the DP performance out of these Wx100 cards while the RX480 Polaris chip should only be good for SP?
Computer ID 10015216
RAMA wrote: I thought that
)
I think there is some DP but the majority is SP. The RX-480 i think is faster on SP?
https://einsteinathome.org/goto/comment/152643
and here
https://einsteinathome.org/goto/comment/152598
Yeah, that was my experience
)
Yeah, that was my experience too - I don't see any significant improvement on GPUs with good (1/2) FP64 support. Even the v1.18 beta which promised improvement on FP64 didn't meet the expectations. SP code seems to be what matters most, so a Fury X or RX 480 with poor FP64 still perform better than Hawaiis which had good FP64.
I measured 660 secs on RX 480 too, but my CPU was heavy occupied with other CPU tasks. Suspending all other tasks and letting the CPU thread supplying GPU reach max turbo clock can yield ~520 secs there.
So CPU performance still plays a significant role, more significant than good FP64 support.
-----