Einstein@Home GPU Application for ATI/AMD Graphics Cards

[PoppaGeek]

Would love to help. But with the heat in TN my GPUs are collecting dust.

Hopefully there will still be interest once things cool off and I come back!

[stoneageman]

Team now 2nd by RAC

[Snow Crash]

glad to be part of the carnage
Mumak was pushing my number 2 position so I had to push back

[Mumak]

Watch your back

[Snow Crash]

Well played sir

I had a bit of a slip - but no excuses I WILL do better!

[Mumak]

Thanks
OK, let's see I'm still not running at full

[stoneageman]

No.1 Team today & the last week

[Mumak]

Well done

[stoneageman]

New team daily record
Team still #1 RAC
Should reach #30 position in about 10 days.

[Snow Crash]

New team daily record
Team still #1 RAC
Should reach #30 position in about 10 days.

Finito!!!
now just another week +1 for #29

Would love to help. But with the heat in TN my GPUs are collecting dust.

Hopefully there will still be interest once things cool off and I come back!

... back ...

[PoppaGeek]

Thanks SC. Nice to have the GPUs humming along again and seeing some numbers posted.

Slowly ramping up so more to come.

Now lets show some science nerds who Team XS is!



EDIT:

Noticing a few things.

E@H works with Crossfire ON.

Nvidia take 50-70% longer for same work unit.

7870 takes 50% longer than 290x for same WU. 290x has faster bandwidth and memory otherwise the same. IIRC.

Question:

If I have 3 7870s in a system and want to play a game on the first 2 GPUs which are crossfired can I get the app to stop using the first 2 but continue on the third GPU which is NOT in the Crossfire? Looked through settings on Wiki and could not find anything.

AND another EDIT:

So 660 ti has a 192 bit bus, 7870 has 256 bit bus and the 290x has a 512 bit bus.

Perseus times: 660 ti = 25000 seconds, 7870 15000 seconds, 290x 9500 seconds,

I am running 2 WUs on each. Maybe 1 on 660 and 2 on 7870 and 3 on 290? From what I have read E@H uses a lot of bandwidth yes/no? hmmm I think I'll see what happens.

[Mumak]

Yes, BRP tasks require quite lot of memory bandwidth. Also note, that for AMD GPUs (BRP OpenCL version) the tasks consume a lot more CPU resources, so make sure to have at least 1 CPU core free. My experience is that a slower AMD GPU with a faster CPU can achieve better times than a slightly faster GPU with a slower CPU:
Here for example running 1 BRP5 WUs:
Core i7-3820 + HD 7950: 4700 s
AMD FX-8350 + HD 280X (7970): 5000 s

Your WU times are quite high, here are a few that I noted on my GPUs (running 2 WUs/GPU):
660 Ti: 13800 s
R9 280X: 7400 s

Also make sure to check your results - if your machine is getting too many invalids, check your GPUs. There have been several reports (including my own experience) that on some HD 280X (HD 7970) which have a high memory clock you'll need to downclock it. I had to reduce it from 1500 to 1250 MHz...

[PoppaGeek]

Thanks Mumak. Only one invalid so far and it was the 270x. I forgot to leave a core free on the 660 ti. Hopefully freeing one will help times.

I have no idea why but I said in the other post I had a 290x it is a 270x. I am old and semi-senile. Yep I'll go with that.

From everything I have read it seems 2 WUs at a time is optimal on most all GPUs and since all the AMDs are on a machine with 2 cards I'll leave at 2 as the PCI bus has it's limitations.

The 660 ti always did pretty decent on GPUGrid, may move it over there.

[jeanguy2]

Hi guys,

I am just in the process of setting up the first rig an i7-3930K with 2 X 7970 in crossfire for Einstein GPU crunching. Rather than going trial and error and wasting valuable crunch time can one of you guys post the best recommended settings to use for the "Einstein@Home preferences"?

I am planning to run 2 X Einstein BRP5 instances per GPU with 1 CPU core assigned to each instances. Do they have to be full cpu core or are we talking threads here? I will the use the remaining cores/threads on WCG crunching FAH

Here is the app_config I plan to use:

<app_config>
<app>
<name>einsteinbinary_BRP5</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>1.0</cpu_usage>
</gpu_versions>
</app>
</app_config>

Any other hint most welcome, Thanks

[jeanguy2]

Hi Guys,

Here is a little update I got the above mentioned rig running with 10 FAH CPU WUs and 4 GPU instances of Einstein@home, 2 instances per GPU. They run as (0.5CPUs + 0.5 ATI GPUs). I had to mod the app_config file as follow to achieve this:

<app_config>
<app>
<name>einsteinbinary_BRP5</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.5</cpu_usage>
</gpu_versions>
</app>
</app_config>

Right now I have a mixture of BRP 4 and BRP 5. I have modified the Einstein@home profile to receive only BRP5 from now on as suggested by PG. The BRP5 appear to be headed for a 2:45 hour completion while the BRP4 are headed for 1:05 hour completion.

If everything is OK in the morning I will setup another rig with a 2600K and 2 X 7970 to crunch for this Einstein GPU project.

[PoppaGeek]

Welcome Jeanguy2.

E@H tasks do better with 1 free core. An Intel user will need to tell you if a HT thread is sufficient as I do not know but think it probably is. The decision comes down to who do you want to hurt. WCG or E@H. I wrestle with it way too much. Crunching cancer work units are important to me but E@H is my fun project. Love the science and the project and the E@H forums. Great people, interesting threads and links. Lots of good reading. In the past I have given preference to WCG but this time I am giving E@H a bit more. I have experimented a bit with .5 CPU and 1 CPU and it makes a difference depending on the GPU, CPU and WU.

My guess is a 7970 will be a hungry beast and more CPU will very much make a difference. I would also guess that the Perseus BRP5 is the best work unit for the 7970 points wise.

While there are people running more than 2 work units on a GPU I have seen little to no evidence that more than 2 increases RAC.

[PoppaGeek]

I started a thread at E@H that might be of interest to owners of older Nvidia GPUs.

GTX 460 better times than GTX 660 TI

[Mumak]

For the 7970 I would definitively suggest
<cpu_usage>1.0</cpu_usage>

Also as I mentioned before, make sure to check if you get invalids.. If that's the case, you'll need to underclock the GPU memory.

[PoppaGeek]

Another thing about E@H. Credit for work units can take awhile to show up. The turn around IIRC is 2 weeks. If you are not getting invalids then the points will climb after 2 or 3 days usually.

[Snow Crash]

BRP5 will get the best points
BRP4 will get second best points
BRP5 and BRP4 at the same time run horribly (long discussion on Einstein and I have also observed this myself).

1 Intel hyper-thread free per WU is sufficient (you can squeak 1%-2% more utilization out of the GUO with more but it's not worth it)

BRP tasks (both 4 and 5) are very PCI hungry, I'm guessing that with 2 7970 you will be bottle necked by PCI and not GPU or CPU.
Nvidia cards that do not increase performance above 2 (tested on 480, 660Ti and 670)
AMD cards need to be tested for best efficiency, for example my 7950 loves 4 WUs but my 7850 does best running 2 concurrent.

Einstein GPU tasks are about as hard on our equipment as it comes. For NVida gear it is about the same as GPUGrid, for AMD - I've never run another project that is qute as demanding so any OC you've base lined on other projects is to high.
GPU speed is more important than memory but once you've arrived at your best GPU / power ramp, memory does help quite a bit.

Keep it cool and rock on

[stoneageman]

Wish they'd get their finger out & release the next batch of GW tasks

[jeanguy2]

Thanks for your replies PG & Mumak,

So far so good, 24 Einstein WUs have now validated. 21 x BRP4 were granted a flat credit 1000. Their running time average 48 minutes. The validated 3 x BRP5 WUs were granted a flat credit of 3300 and ran for an average of 3 hours & 15 Mins. There are 49 other Einstein WUs sitting in the PD box.There are no invalids or errors reported so far. The 7970s are running @ stock speed of 1010/1250 with still a 0.5 CPU allocation for each GPU instances. So far it appears that a flat credit of 1000 per crunching hour applies to the BRP4 & 5 that I completed. BRP4 complete quicker so they maybe more efficient on this rig.

Enough talking, time to setup the second rig to get a couple more 7970 crunching for the XS Team.

UPDATE: 2nd rig now operational with 2600K CPU and 2 x 7970s with 4 Einstein BRP5 WUs crunching away. (2 WUs per 7970) I will likely add a couple of 7950 tied to a 3770K CPU this weekend.

[PoppaGeek]

Milestones Today
stoneageman passes 190,000,000
Raqia passes 800,000
jeanguy2 passes 100,000

[stoneageman]

If my output keeps as it is, in less than two weeks I should gain 8 places The team is looking in good shape

[jeanguy2]

If my output keeps as it is, in less than two weeks I should gain 8 places The team is looking in good shape

Great work Sam... It also looks like you are going to take #2 spot for your country from "chase1902" leaving only a chap named "Gavin" in your crosshair for the #1 in country spot. You are definitely operating in the heavy hitters daily credit area. Keep at it Sam...