ATI 7970 and Intel GPU = bad performance
19 Dec 2013 20:33:46 UTC
Topic 197316
(moderation:
I enabled the Intel GPU (i5-4440) a week or two ago. I change the BIOS to enable it as the initial display and got BOINC running with it.
However, I discovered that the overall performance started to drop off. I removed the 'Priority by eFMer' for the Intel app, but still kept falling.
I then disable the Intel tasks, but still no go.
Lastly, I went back into BIOS and disabled the GPU all together and removed driver, but still very poor performance.
The computer http://einsteinathome.org/host/6684790 has gone from Ave. RAC of around 111k to now running at around 70k and falling. The run times for the tasks is what is causing the problem, but I cannot understand why they would all of a sudden take 50% longer.
The only thing I changes as to install/enable the Intel GPU, but I have since completely removed this.
I have Priority running for the GPU tasks as well as a very small overclocking. I also upgraded the drivers to the latest beta drivers but still no change.
ps. I have another host running and older nVidia GTX 570 and Intel HD 4600, which looks to be running happily together, http://einsteinathome.org/host/9615019.
Any suggestions, performance currently running at around 1/2 of best performance before the changes.
Arvid
Arvid Almstrom
ATI 7970 and Intel GPU = bad performance
)
Try reducing the number of CPU tasks running on the computer - first by one, then by two, and report back, please.
I run host 8864187 (i5-4570 with HD 4600). Speed when all four traditional 'CPU' cores is greatly reduced, but it's OK at 75% loading.
Hi Richard I am only
)
Hi Richard
I am only running 2 x CPU and 4 x GPU (0.25 CPU and 0.25 GPU). The system is running at about 60-65% CPU usage.
This is the same setup that I had a couple of weeks ago when all was well.
If you want I can disable all CPU but I don't know if that is the problem.
Many thanks
Arvid Almstrom
OK, my host has no separate
)
OK, my host has no separate discrete GPU, so my experience will not be directly comparable.
I see you've run every class of application available:
'Gamma-ray pulsar search #2' and 'Gravitational Wave S6 Directed Search (CasA)' on CPUs
'Binary Radio Pulsar Search (Arecibo, GPU)' and 'Binary Radio Pulsar Search (Perseus Arm Survey)' with opencl-ati
'Binary Radio Pulsar Search (Arecibo)' with opencl-intel_gpu - now discontinued.
When you say "The run times for the tasks is what is causing the problem", which tasks are those - all types, or just some? If some, which?
Please see the list below of
)
Please see the list below of average processing times for the daily WU's for the the ATI/AMD GPU are listed below.
The BRP5 WU's have gone from around 10,000-11,000 to 14,000-16,000.
Likewise the BRP4G WU's have gone from the low 3,000 to 5,000.
DateTime BRP4G-opencl-ati BRP5-opencl-ati
9/11/2013 00:00:06 3317 10575
10/11/2013 07:11:22 3167 10818
11/11/2013 00:00:05 3123 10659
12/11/2013 00:00:05 3170 10606
16/11/2013 11:33:51 3206 10172
17/11/2013 00:00:06 3173 10751
18/11/2013 00:00:05 3231 10904
19/11/2013 00:00:07 3317 10708
20/11/2013 00:00:06 3219 10809
21/11/2013 00:00:05 3202 10799
22/11/2013 00:00:05 3147 10646
23/11/2013 00:00:08 3161 10787
24/11/2013 00:00:06 3264 10756
25/11/2013 00:00:09 3178 10910
26/11/2013 00:00:08 2940 10633
27/11/2013 00:00:05 3267 10643
28/11/2013 00:00:06 3196 10609
29/11/2013 00:00:06 3174 10800
30/11/2013 00:00:05 3498 11485
1/12/2013 00:00:06 3962 12903
2/12/2013 00:00:06 4110 13405
3/12/2013 00:00:07 3929 13859
5/12/2013 00:00:11 4012 13945
6/12/2013 00:00:09 4052 14692
7/12/2013 00:00:08 3813 12801
8/12/2013 00:00:08 3722 12311
9/12/2013 00:00:09 4018 12093
10/12/2013 00:00:10 3425 11885
11/12/2013 00:00:06 3636 12633
13/12/2013 00:00:10 4738 15763
14/12/2013 00:00:10 4983 15536
17/12/2013 00:00:16 5052 16313
19/12/2013 07:56:41 5052 16460
20/12/2013 00:00:31 4960 11918
Arvid Almstrom
So it's the opencl-ati tasks
)
So it's the opencl-ati tasks in both cases?
I wondered about that. Maybe the common OpenCL components are different between the two driver installations, and the second (Intel?) replaced the first (ATI?) set with something slower?
RE: ... Maybe the common
)
I think you are exactly on the money. I don't run any Windows machines so I have no experience so please take these comments with a huge grain of salt.
I have a number of Linux installations using AMD GPUs - 7770s and 7850s. The Linux distro I prefer to use has reasonably up-to-date catalyst drivers in the repo but if I try to crunch with them the performance is woeful. I had previously solved the problem by using openSUSE on those machines. This gave me full performance.
Recently I've been experimenting with a new install using my preferred distro. It's actually the host I've just documented in this message in another thread. Here's some detail about what I did to get full performance.
After installing and updating the preferred OS (PCLinuxOS), I completely removed the repo supplied driver (supposedly Catalyst 13.4). I went to the AMD website and downloaded a fresh copy of Catalyst 13.4 for linux and the latest AMD APP SDK (version 2.9). I read somewhere on the website that the GPU runtime was now part of the catalyst driver suite and that it was important to install the Catalyst driver first before the SDK otherwise you would get the CPU runtime instead of the GPU runtime - or something along those lines.
On a hunch that my problem of slow running might be something to do with the wrong runtime, I installed the AMD supplied driver and then started BOINC to be greeted with the missing GPU message. So I installed the SDK and now the GPU was properly identified and was running at full performance. I don't know why I needed to install both (since I wasn't going to be developing OpenCL apps), but who cares, the thing is running beautifully :-).
I just wonder if Arvid has now got the the wrong runtime and perhaps all he needs to do is remove the current drivers and install fresh copies from the AMD website. I notice there is now a 13.12 WHQL driver available. If things don't work properly after a driver install, perhaps installing the Windows version of the SDK might get things going.
Cheers,
Gary.
Just a wild guess but since
)
Just a wild guess but since you are living "down-under", was there a noticeable increase of ambient temperature? Is the GPU running at full clock rate all the time or is it throttling down (for whatever reasons).
Cheers
HB
I've got an AMD 7970 GPU
)
I've got an AMD 7970 GPU working on MB with PCIE2.0 only support, so I can tell you, that your host's runtimes are quite similar to GPU working at PCIE2.0 16x. Are you positive that your 7970 works at full speed PCIE3.0?
Thank you for all your
)
Thank you for all your feedback and sorry for not getting back here sooner, but I have been away for a few days.
I have, after some issues, re-installed things on the computer but still had issues.
Until I discovered the speed that the card was running at. I had earlier overclocked the card by about 5% to get a little more performance out of it and thought nothing of it.
However, when I had tried everything else I thought I would reset the card to the defaults and found that after the installation of the Intel GPU drivers, the ATI card had been down-clocked by some process. It was running at 500MHz and the RAM was running 2,800MHz. These should be 700MHz and 4,700MHz.
Once I reset these to the defaults and added 5%, it now looks like things are back to somewhere like they should be. I will leave it for a few days and see.
I just cannot understand how something in the system would have down-clocked the GPU! I know that if the card gets too hot it would automatically throttle the card, but only while conditions are outside the specs for the card and should then come back to the normal speeds. I can only think that the process of setting the Intel GPU as the default in BIOS, to enable it in Windows, that the BIOS or something in Windows would have slowed the ATI card down.
I am sorry for something that simple to cause this much problem, but I just didn't look at the actual speed of the GPU as I had set them at a small overclocking and never touched them again.
Thank you all for taking the time to give me a help on this issue.
Yes it has been hot down here and I have had to shut down the computers a couple of times, but the run-times, should still be the same as long as the cards are within specs.
Gary, I will have a good read of your post about the issues with drivers vs. SDK setup and make sure. However, I am running Windows and don't think that is the problem here, but you never know.
ps. I am just amazed of the performance of the ATI/AMD cards vs. the nVidia cards. I have only ever bought nVidia cards in the past, but thought I would try an ATI/AMD card. This computer with just one 7970 will do just about as much as my dual GTX 780!
I guess that I will stay with ATI in the future, at least for E@H work, as they are more than twice the speed while using opencl.
Many thanks
Arvid
Arvid Almstrom
RE: I guess that I will
)
Indeed, 8 out of the top 10 hosts of E@H are AMD/OpenCL driven. This was somewhat surprising to me, as the FFT lib that we use for the OpenCL apps is a) quite old and b) originally optimized for NVIDIA GPUs, while the FFT lib we use for the CUDA apps is NVIDIA's own cuFFT (but also a quite old version). We will soon look at the clFFT lib, the open source spin-off from AMD's APPML package, as an alternative for OpenCL apps, and newer cuFFT libs as alternatives for the CUDA apps.
Cheers
HB