Question Does anyone have a 7900 XTX?

[Skillz]

Looking for someone with a 7900 XTX and a 7900 XT would also help since both are similar, but mainly want to know the 7900 XTX.

Both of those cards have decent amounts of FP64 commute which should make them run a BOINC project called Milkyway@home really well. I'm curious to know how well they actually perform when running this project.

If you have one of these cards, would you be so kind to install BOINC, join Milkyway@home, run a few GPU Separation tasks?
In your account on MW@home, make sure you uncheck CPU so it doesn't run CPU tasks. Also uncheck n-body sub project so it wont run that sub project (its CPU only anyway)

I would like details on how quickly it completes 1 task.
Then I want details on how fast it completes, 2, 3, 4, 5, 6, 7, etc... tasks simultaneously until we get to a point of diminishing returns or it uses up all the VRAM.

Typically these tasks take less than a minute to run on a good GPU and up to a few minutes on a not-so-good GPU. So the testing time shouldn't take up too much of your time.

If you need help with installing BOINC, joining MW@Home, attaching the project, configuring BOINC to run multiple tasks on the same GPU or any other help with getting this working I can help you.

 

[BGod]

I know it's not the project you asked about, but maybe the Dnet OCL client will give you an idea:
AMD RX 7900 XTX: 27,105,601,713 keys/s

 

[Skillz]

I know it's not the project you asked about, but maybe the Dnet OCL client will give you an idea:
AMD RX 7900 XTX: 27,105,601,713 keys/s

Thank you, but it doesn't.

Milkyway GPU Separation tasks makes heavy use of FP64 commute which is pretty unique. Not a lot of software utilize FP64 which is why most gaming cards are pretty slow in that department.

I also have nothing to compare those numbers to.

May I ask what Dnet OCL is exactly?

 

[Blender111]

I think I finally got it working w/ just the GPU and it takes 35-40 seconds with my 7900XTX, probably averaging 37-38. That's for a single task, I couldn't figure out how to get it to run more than 1 at the same time.

 

[BGod]

Thank you, but it doesn't.

Milkyway GPU Separation tasks makes heavy use of FP64 commute which is pretty unique. Not a lot of software utilize FP64 which is why most gaming cards are pretty slow in that department.

I also have nothing to compare those numbers to.

May I ask what Dnet OCL is exactly?

The Stream client got broken some where along the line with an AMD driver update, so now I'm stuck running the Open CL version.

 

[Skillz]

The Stream client got broken some where along the line with an AMD driver update, so now I'm stuck running the Open CL version.

I still don't know what that is.

 

[BGod]

I still don't know what that is.

I'm one of the old school RC5 members. I did the RC5-64 bit contest mostly for AnandTech, but switched to The Old Republic before it ended. Worked on the RC5-72 bit contest until I crunched 10 million blocks and gave it up for a number of years. Now that the GPUs have become so powerful, I had to dive back in.

 

[waffleironhead]

I still don't know what that is.

distributed.net: Project RC5

its the same project as moowrap on boinc.

Moo! Wrapper (moowrap.net)

 

[Skillz]

Ahh okay. Moo! Wrapper doesn't utilize FP64 compute, so its data isn't relevant to what it could do in Milkyway@home. But thanks for the insight and information.

 

[waffleironhead]

Does any other project use fp64?
I have my a12-9800 on milkyway as the IGP has 1/2 DP. Its about the only good thing going for that particular APU.

 

[Skillz]

Does any other project use fp64?
I have my a12-9800 on milkyway as the IGP has 1/2 DP. Its about the only good thing going for that particular APU.

Milkyway is the only project I know of that utilizes it. Some PG sub projects used to use it, I think, but they've since updated them to not use it anymore.

 

[Shmee]

If I can get one I can give MW@home a try for you. Hoping to pick up a Sapphire Nitro 7900 XTX later today, if I can grab one in stock.

 

[Shmee]

I downloaded BOINC and I am running MW@home now on the card. I admit I cannot figure out how to disable CPU tasks. Also, is there a way to tell BOINC to finish current tasks, and then suspend work? I think F@H could do that.

Anyway, I am already logged in to the team Anandtech, so we should be getting points.

 

[Skillz]

@IEC would you be willing to test your 7900 XT on Milkyway@home for me?

 

[IEC]

@IEC would you be willing to test your 7900 XT on Milkyway@home for me?

 

[IEC]

At least on paper similar DP performance to my Radeon VII which is currently sitting unused.

 

[Skillz]

Do you know how to run multiple tasks at once using BOINC? I'd like to see the completion time with many tasks running at once.

For example. I run 8 tasks at a time on my P100s each with a minute to a minute and a half completion.

Can you do some tests on that to see how fast it completes 2, 3, 4, 5, 6, etc... tasks at a time until it degrades performance?

 

[Skillz]

I downloaded BOINC and I am running MW@home now on the card. I admit I cannot figure out how to disable CPU tasks. Also, is there a way to tell BOINC to finish current tasks, and then suspend work? I think F@H could do that.

Anyway, I am already logged in to the team Anandtech, so we should be getting points.

I totally missed your post. You can set "No new work" in the BOINC manager and BOINC will not request new tasks when it completes the current queue of tasks. You can set the queue low so it wont download a lot of tasks at once.

 

[Shmee]

I totally missed your post. You can set "No new work" in the BOINC manager and BOINC will not request new tasks when it completes the current queue of tasks. You can set the queue low so it wont download a lot of tasks at once.

Thanks, I will give another try setting that up, and try to run it GPU only. How is the best way to report how well the card does in? Should I take screenshots of BOINC after tasks are completed, or should I link to my account or something on the MW@Home webpage?

 

[Skillz]

You can select "GPU Only" on the Milkyway@home preferences. Just uncheck the CPU setting so it wont send you CPU work.

The best way is to well, monitor it yourself. Since you'll be running different settings it wont separate the different settings on the MW web site when it reports the tasks. Personally, when I am doing this I use a program called BoincTasks. It shows a history tab of all the completed work units and the time it took to complete them.

When I make changes, I write down the last WU reported in the history tab and that way I know everything before that mark is the new settings.

However, with Milkyway@home the tasks run so fast on decent cards (couple minutes at most) that I just watch the tasks live to see what their completion times are.

 

[Shmee]

Ok, well I was watching a few GPU tasks, on average most would complete in about 30 seconds.

 

[Skillz]

Ok, well I was watching a few GPU tasks, on average most would complete in about 30 seconds.

Was that running just one task at a time?

 

[Shmee]

Was that running just one task at a time?

I think it was just doing one GPU task at a time, but not certain. I know it was also doing CPU tasks. I will fiddle with it some more, and see if I can turn off any CPU tasks, and try turning on multiple GPU tasks, if I can figure it out lol.

 

[Skillz]

To run only GPU tasks uncheck CPU here: https://milkyway.cs.rpi.edu/milkyway/prefs.php?subset=project (you will need to login)



This will only give you GPU work on either AMD/ATI or NVIDIA GPUs.

To run more than one task concurrently on the GPU make a file named app_config.xml and place it in the C:\ProgramData\BOINC\projects\milkyway.cs.rpi.edu_milkyway\ directory (If running Windows; if Linux it'll determine where you installed it usually /var/lib/boinc)

Inside the app_config.xml file add this.

<app_config>
<app>
<name>milkyway</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.05</cpu_usage>
</gpu_versions>
</app>
</app_config>

Take a note of this line:
<gpu_usage>0.5</gpu_usage>

This will allow two tasks to run at once on the GPU.
<gpu_usage>0.33</gpu_usage> - 3 tasks at once
<gpu_usage>0.25</gpu_usage> - 4 tasks at once
<gpu_usage>0.20</gpu_usage> - 5 tasks at once
<gpu_usage>0.16</gpu_usage> - 6 tasks at once
<gpu_usage>0.14</gpu_usage> - 7 tasks at once
<gpu_usage>0.12</gpu_usage> - 8 tasks at once

VRAM might start becoming an issue with more than 8 tasks, but you can continue that formula to get 9, 10, 11+ tasks to run at once.

You MUST restart the BOINC client for settings to take effect.

The CPU usage setting doesn't really do much in this case. Just setting it to .05 just tells BOINC that this task isn't using the CPU and the CPU is free to do whatever it wants. Since you wont be running CPU tasks during the testing it doesn't matter.