Excellent, it's working.
That's it for tonight, thanks for your help.

Matt

Well, finished, but GPU load was very very low (time to time <20%, but mostly no GPU load, only CPU load).
I must re-attach the project to get a new task. Could you add a profile selections for ATI/AMD GPU?

Excellent, it's working.
That's it for tonight, thanks for your help.

Matt

You are welcome.

I got my first task for my 5770 but it errored out. :(

30/01/2015 12:01:22 PM | GPUGRID | Finished download of mdrun-912.exe
30/01/2015 12:01:22 PM | GPUGRID | [error] Signature verification failed for mdrun-912.exe
30/01/2015 12:01:22 PM | GPUGRID | [error] Checksum or signature error for mdrun-912.exe
30/01/2015 12:02:57 PM | GPUGRID | Sending scheduler request: To report completed tasks.

Finally got a task to run but now there are 2 problems.

19252 GPUGRID 1/29/2015 9:11:39 PM Starting task 196-MJHARVEY_AMDDHFR3-0-1-RND2231_5

Problem #1. There is almost no work being done on the GPU. Only an occasional spike to 12% load that lasts about 1 second. CPU load is 99%. Looks like a CPU app to me.
Problem #2. Estimated run time is 55 DAYS. (1336 hours) 0.028% complete after 30 minutes. R9 270X

I'll let it run over night but I don't hold out much hope.

http://www.gpugrid.net/result.php?resultid=13763843

http://www.gpugrid.net/result.php?resultid=13763875
First Task ready. But as was said, only liitle gpu work, mostly cpu. I set 100% cpu time for gpugrid and the wu run with all cpu-cores it could get( as mt predicted)
The max gpu was around 80% for 0,3 sec

I tried to download amd wu's.
Message says: 30.01.2015 09:30:34 | GPUGRID | Not requesting tasks: don't need (CPU: job cache full; NVIDIA GPU: job cache full; AMD/ATI GPU: ; Intel GPU: job cache full)

I have no amd wu eunning, only a nvidia wu and a intel wu.

Seems that the gpu detection does not work correct. Works fine with POEM.

Refers to this pc http://www.gpugrid.net/show_host_detail.php?hostid=187043

When setting the nVidia wu to hold and manually update GPUGRID the sceduler asks for gpu wu's for nVidia and for AMD. Returned message says: no work available. nVidia wu's set to run only short runs.

Alexander

:-)
Run time 254.09
CPU time 838.10
-----------------
score 6,025.54

http://www.gpugrid.net/results.php?hostid=197040

The AMD application has reached a working state.

To access it you will need to change your preferences to
* enable "Accept Test work"
* enable the application "Molecular Dynamics on AMD GPUs"
* enable "Use CPU"

You will also need a 64bit Operating System, Windows 7 or higher.

Unlike the application for NVIDIA cards, this one requires as much CPU as you can assign it.[1]

For technical reasons BOINC considers this application a CPU only app - it won't show up as requiring a GPU. Despite that, you will only receive work if the client reports that the machine has an AMD GPU installed. The WU group name is AMDTEST1.

I am putting through a lot of short test WUs to start to gauge performance estimates and failure modes.


Known issues are:
* Checkpoint restart not working
* BOINC progress reporting not working
* If GPU is incompatible (HD4000/5000), will fall back to CPU execution
* No GPU assignment - will always run on first GPU in system (beware if you have mixed AMD/NVIDIA system)

Matt

[1] It's based on the old - now deprecated - CPU app, but uses a GPU for additional acceleration.

[1] It's based on the old - now deprecated - CPU app, but uses a GPU for additional acceleration.

This makes it very much a niche app, and people running it need to watch out for unexpected behaviour.

In particular, I would expect that most BOINC users who have AMD/ATI GPUs will already be attached to other BOINC projects which have fully-specified AMD/ATI applications. In this mode of operation, BOINC will schedule both projects to use the same device at the same time - that may have an impact on your performance statistics.

In particular, I would expect that most BOINC users who have AMD/ATI GPUs will already be attached to other BOINC projects which have fully-specified AMD/ATI applications. In this mode of operation, BOINC will schedule both projects to use the same device at the same time - that may have an impact on your performance statistics.

Yes, I'm aware of the limitations but it's the best we can do given our current server software.
The current situation is acceptable for now, while we are still in the testing phase.

I'm only getting one task and won't get another one unless I detach and reattach. Also getting this when trying to get new work.

1/30/2015 6:10:49 AM | GPUGRID | No tasks are available for OpenCL for AMD GPUs

I'm only getting one task and won't get another one unless I detach and reattach. Also getting this when trying to get new work.

+1

The AMD application has reached a working state.

To access it you will need to change your preferences to
* enable "Accept Test work"
* enable the application "Molecular Dynamics on AMD GPUs"
* enable "Use CPU"

Doesn't work for me. I have tried all possible combinations of checkboxes without any success. The only way to get a task is re-attaching the project.

30/01/2015 13:11:18 | GPUGRID | Requesting new tasks for CPU
30/01/2015 13:11:24 | GPUGRID | Scheduler request completed: got 0 new tasks
30/01/2015 13:11:24 | GPUGRID | No tasks sent
30/01/2015 13:11:24 | GPUGRID | No tasks are available for OpenCL for AMD GPUs
30/01/2015 13:11:52 | GPUGRID | update requested by user
30/01/2015 13:12:35 | GPUGRID | Sending scheduler request: Requested by user.
30/01/2015 13:12:35 | GPUGRID | Requesting new tasks for CPU
30/01/2015 13:12:38 | GPUGRID | Scheduler request completed: got 0 new tasks
30/01/2015 13:12:38 | GPUGRID | No tasks sent
30/01/2015 13:12:38 | GPUGRID | No tasks are available for OpenCL for AMD GPUs
30/01/2015 13:12:50 | GPUGRID | Resetting project
30/01/2015 13:12:50 | GPUGRID | Detaching from project
30/01/2015 13:13:01 | | Fetching configuration file from http://www.gpugrid.net/get_project_config.php
30/01/2015 13:13:18 | GPUGRID | Master file download succeeded
30/01/2015 13:13:23 | GPUGRID | Sending scheduler request: Project initialization.
30/01/2015 13:13:23 | GPUGRID | Requesting new tasks for CPU and ATI
30/01/2015 13:13:26 | GPUGRID | Scheduler request completed: got 1 new tasks

The AMD application has reached a working state.

Are you sure? My AMD R9 290 says the opposite.
http://www.gpugrid.net/result.php?resultid=13763788

If GPU is incompatible (HD4000/5000), will fall back to CPU execution

I got an AMD Mobility 5650 in my Laptop and it's working fine (~87 % GPU load).

In particular, I would expect that most BOINC users who have AMD/ATI GPUs will already be attached to other BOINC projects which have fully-specified AMD/ATI applications. In this mode of operation, BOINC will schedule both projects to use the same device at the same time - that may have an impact on your performance statistics.

Yes, I'm aware of the limitations but it's the best we can do given our current server software.
The current situation is acceptable for now, while we are still in the testing phase.

I appreciate the effort to get an AMD app to run on GPUGrid but I'm going to pass on this one. IMHO an MT GPU app that requires so much CPU power is no more than a glorified CPU app. Good luck. I'll keep an eye on this and if it ever develops into a true GPU app I'll try again.

I can not get a couple of days jobs from amd beta..

AMD cards are very inefficient. to my 3 7970 1200I corsair must watt at full throttle power supply. include cpu
maybe people want to help their suppliers of electricity.-) dumb
but the path to a stable amd applications will be long and seamless.
Now amd it was cheaper to 290x card which I returned after 14 days due to black screeen problem. and still be switched to throtle mod. absolutely useless. It was the time when she was a complete novelty and become full price.
AMD card can now buy just a fool like me.

Got one running now. I get it, this is the old multi-threaded CPU app that has been enhanced to share some of the processing onto the GPU.

When the task is running, BOINC manager shows that it is using 1 CPUs + 1 ATI/AMD GPUs. However when I look at the Stderr.txt, it shows that the task is using 16 threads (all that is available on this machine). And I can confirm that task manager showed 100% cpu utilization.

BOINC wrapper for GROMACS.
Arg 0	[projects/www.gpugrid.net/mdrun-915.exe]
Arg 1	[--nthreads]
Arg 2	[16]
BOINC running with [16] threads

And at that, it only uses the GPU part of the time. When the GPU is being used, it is being used at 90% (on an AMD 6950). Then it goes for a minute or so of 0% being used.

Task ran for ~20 minutes and now shows as completed and validated.

Edit:

When I add the GPUGRID project and ask for "OpenCL for AMD GPUs" tasks, BOINC manager requests tasks for CPU and AMD GPU. It gets one task. After it gets that one task, whenever it requests new tasks, it only requests tasks for CPU. Then it never gets another task. The only way to get another task is to detach and re-attach, then BOINC Manager will request tasks for CPU and AMD GPU.

Got one running now. I get it, this is the old multi-threaded CPU app that has been enhanced to share some of the processing onto the GPU.

More or less, yes.

When the task is running, BOINC manager shows that it is using 1 CPUs + 1 ATI/AMD GPUs. However when I look at the Stderr.txt, it shows that the task is using 16 threads (all that is available on this machine). And I can confirm that task manager showed 100% cpu utilization.

It should have the same behaviour as the old CPU app, using whatever CPU you have allocated to the project in the preferences. (I can see that's working for other clients)

And at that, it only uses the GPU part of the time. When the GPU is being used, it is being used at 90% (on an AMD 6950). Then it goes for a minute or so of 0% being used.

Shouldn't have a duty cycle like that. The peak load is very likely to be below 100% until we've tuned the CPU/GPU balance, but it shouldn't spend a long time idle.

When I add the GPUGRID project and ask for "OpenCL for AMD GPUs" tasks, BOINC manager requests tasks for CPU and AMD GPU. It gets one task. After it gets that one task, whenever it requests new tasks, it only requests tasks for CPU. Then it never gets another task. The only way to get another task is to detach and re-attach, then BOINC Manager will request tasks for CPU and AMD GPU.

Yes, a lot of people have reported that. Could you show me the client output showing that behaviour, please?

Matt

When I add the GPUGRID project and ask for "OpenCL for AMD GPUs" tasks, BOINC manager requests tasks for CPU and AMD GPU. It gets one task. After it gets that one task, whenever it requests new tasks, it only requests tasks for CPU. Then it never gets another task. The only way to get another task is to detach and re-attach, then BOINC Manager will request tasks for CPU and AMD GPU.

Yes, a lot of people have reported that. Could you show me the client output showing that behaviour, please?

Matt

From captainjack's description, it's probably the line

<no_ati_apps>1</no_ati_apps>

which I've copied from your sched_reply_www.gpugrid.net.xml file (below the project preferences block, and just before the <gui_urls>).

I don't have any computers with an AMD/ATI card, but that probably translates into a line

<no_rsc_apps>ATI</no_rsc_apps>

in the <project> section of client_state.xml, and having the effect of preventing BOINC making any further requests for AMD/ATI work.

[I have <no_rsc_apps>NVIDIA</no_rsc_apps> from another project, and that has the effect of making

30/01/2015 17:10:55 | LHC@home 1.0 | [work_fetch] share 0.000 no applications

appear in the <work_fetch_debug> event logs]

Someone could try searching for <no_rsc_apps>ATI</no_rsc_apps> in client_state after processing their first test app, and see if removing it allows a second bite of the cherry.