If you don't have the right drivers installed for the graphics card you also won't get work. So some of the minimums are to have the right card, the later version of drivers installed and the right version of BOINC (6.4.2?) or later.

As best as I have been able to tell, if you get the CUDA message in BOINC on start up you are good to go. However, there also may be limits on the VRAM as in if you don't have enough you also won't get work.

All this is some what of a speculation on my part in that I have only started looking at this since the announcement so ... but, I have done two tasks for GPU-Grid successfully so far, with lots of failures (5-10 by now) and mostly get a no work available message.

The GPU-G people say to stick with the earlier version of BOINC that is CUDA capable (one I *DON'T* have) in that they suspect 6.4.5 has some sort of problem. I have gotten some work from them, but not a lot so it is too early for me to tell.

In that I am not that into SaH anymore, I have no experience with what is going on here. I just started doing a little more work though I am waiting for my numbers to settle since I added one new system and took two off and changed the distribution of projects and reallocated resource shares amongst my project suite...

{edit}It is also not clear if having two GPU, or more, is the "best" way to actually use this capability so as to have several GPU working off of one CPU that is feeding it work.

Secondly, you will not see the GPU task in Task MGR as that only monitors the CPU usage and not the GPU at all ...

Last, the GPU does many more tasks in parallel so that if you can paralelize the operations then you can see incredible increases in throughput.
{/edit}

---

Hi Paul :)

---

mambo

Hi Paul :)

Hi Francis ...

I just hope I am not going to regret speaking again ...

---

Well the release of the CUDA-enabled client is rocking the proverbial boat a bit, changes to established systems do that, but so far people seem to be in good spirits.

Me, I'm just trying to figure out if I'll be able to crunch SETI on the CPUs and GPU at the same time or if now would be a good time to look into the Einstein project to keep the CPUs warm and let the GPU find ET or if I should wait it out a bit. I'm gunna be keeping an eye on this here thread.

Either way it was nice being able to say Hi :).

---

mambo

Im wondering if i should abandon AK8 and go stock, because now it doesnt seem like its running any CUDA.

Even though i get this message when i fire up boinc client:

CUDA Devices found
Coprocessor: GeForce 8800 GTS (1)

So if i run stock application the cuda would kick in i suppose?

I have exactly the same problem. I also have a 8800GTS (with 320MB memory) and I don't get a single workunit for my Graphics card. Could it be the same problem as there is with GPUgrid: namely that these cards do support CUDA but not the more recent CUDA with which the application is compiled?

Well the problem with the 8-series is that some of them will do CUDA and others will not. Yes they have the same model number and specs, but it depends on the version of the core they are using. My 8800GT was bought about six months ago and it is the G92 core, which does CUDA. Apparently 8800GTs that were bought over a year ago (don't know the exact timeline) perform nearly identically when put to work with a game or other means of benchmarking, but are not CUDA-capable.

Well, my 8800GTS does CUDA, I also get the CUDA message in BOINC, the problem with GPUgrid is that it can't do the CUDA applications which are built with the newer compiler. So the real question is: can all CUDA capable graphic cards do the SETI applications(as is the case with folding@home e.g. as I have experienced) or do you need to have a graphics card that can deal with the more recent CUDA code.

i think if you install boinc 6.2 and 6.4 in two folders then run them both with the 6.2 set to not use one core that you could have it use the cpu and gpu at the same time have not tested this will some else?

i think if you install boinc 6.2 and 6.4 in two folders then run them both with the 6.2 set to not use one core that you could have it use the cpu and gpu at the same time have not tested this will some else?

George,

I think that you are going to have one CPU tied to feeding the GPU regardless. This is one of the not clear issues. *IF* this is the case, and if your processing is just as fast with the CPU alone, then, CUDA is not a good choice.

*IF* you have, on the other hand, a gaming rig with multiple high end graphics cards it is *POSSIBLE* that you would have one CPU monitoring and feeding the several GPUs (whether this would be to run several tasks at the same time or increase the parallelization of one task is also unknown at this time). Is this scenerio you would have "lost" one CPU and gained one or more GPUs. My latest rig can have up to 3 video cards installed (with another question about bus speeds in that the manual implies with 3 cards only one runs at 16x, but with one or two the cards use a bus speed of 16x, and how this affects processing speed is another question).

For me to satisfy my curiosity and to sort of answer your implied question I would need to buy 3 video cards (one high end and two of the same sort I am using in the new rig) and then mix and match while running work.

My own hope would be that I could "lose" one CPU but gain 3 GPU for a net gain of 2 processors. Of course the next problem is that projects that stand to gain by this are not projects I am that interested in supporting .... :)

Decisions, decisions ...

---

i think if you install boinc 6.2 and 6.4 in two folders then run them both with the 6.2 set to not use one core that you could have it use the cpu and gpu at the same time have not tested this will some else?

Why not try it yourself before suggesting something like that ?
Wouldn't it be cool to post a I know if ... instead of "I think..." ?

...but before you actually try that, let me tell you that i b}know[/b] that it won't work...

---

Join BOINC United now!

i think if you install boinc 6.2 and 6.4 in two folders then run them both ...

Can't be done.

1. BOINC checks if another client is already running and if it is, the 2nd started will give a message about that and not run at all.

2. There's only one registry entry for BOINC, so if you install 6.4 first, then try to install 6.2 in another directory, this install overwrites the registry settings of 6.4, including where the Data directory is ... if you then start 6.4, it will run with all the preferences of 6.2, in the directory of 6.2

---

Can't be done.

1. BOINC checks if another client is already running and if it is, the 2nd started will give a message about that and not run at all.

2. There's only one registry entry for BOINC, so if you install 6.4 first, then try to install 6.2 in another directory, this install overwrites the registry settings of 6.4, including where the Data directory is ... if you then start 6.4, it will run with all the preferences of 6.2, in the directory of 6.2

your right i trid it

I think that you are going to have one CPU tied to feeding the GPU regardless. This is one of the not clear issues. *IF* this is the case, and if your processing is just as fast with the CPU alone, then, CUDA is not a good choice.

i have a good gpu 260gtx its a lot faster then just cpu but there is a unused cpu core i would like to use

*IF* you have, on the other hand, a gaming rig with multiple high end graphics cards it is *POSSIBLE* that you would have one CPU monitoring and feeding the several GPUs (whether this would be to run several tasks at the same time or increase the parallelization of one task is also unknown at this time). Is this scenerio you would have "lost" one CPU and gained one or more GPUs. My latest rig can have up to 3 video cards installed (with another question about bus speeds in that the manual implies with 3 cards only one runs at 16x, but with one or two the cards use a bus speed of 16x, and how this affects processing speed is another question).

i have a low end gpu on the motherboard as well and it runs a 2en task i then ask a friend to test some thing and as long as the cards are not in sli they ea run there own task if there in sli there are seen as one bigger gpu and just do one task faster

My own hope would be that I could "lose" one CPU but gain 3 GPU for a net gain of 2 processors. Of course the next problem is that projects that stand to gain by this are not projects I am that interested in supporting .... :)

ea gpu task uses 6% of the cpu
based on a 5400+ amd duel core feeding a 260gtx all stock clock
so one cpu can handle 4 gpus 5 if one is on the motherboard lag may be slow down the task but duel core should do fine


edit

two the point for ea card or pair of cards in sli boinc will runs one task and see it as using .04 cpu and 1 cuda

Just to amuse myself on my computer W2 I just DL 11 tasks and they are running at about 50+ seconds on the clock report but actually take about 7 clock minutes to run per task.

CPU usage is 0.03 and that reduction shows on the Task Manager (GPU Grid was running at 0.90 on the CPU while the tasks were running). Maybe I should canablaize the other GPU card I have and put two in the one machine to see what it would do in that case.

But, I think I will let these run out and then I will see how they fare in validation ... looking at the first halh there is a good mix of pulses and triplets found (no gaussians yet) ... still, no -9 but not matched with a wingman.

Typing going the heck so time to get a nap I guess ...

---

The big problem I have is that if you have a CUDA capable GPU and run SETI from what I understand, you can't prevent it if you have client 6.4.5. The facility to turn it off is coming later according to the CUDA FAQ.

I'd have liked to have this functionality from the start in case something went wrong and I wanted to go hack to the CPU version.

Looks like I'll be reinstalling 6.2.19 until this is sorted

Better install AK8 opr app with app_info.xml. And you will never recive CUDA app if you not want it.

i think if you install boinc 6.2 and 6.4 in two folders then run them both ...

Can't be done.

1. BOINC checks if another client is already running and if it is, the 2nd started will give a message about that and not run at all.

2. There's only one registry entry for BOINC, so if you install 6.4 first, then try to install 6.2 in another directory, this install overwrites the registry settings of 6.4, including where the Data directory is ... if you then start 6.4, it will run with all the preferences of 6.2, in the directory of 6.2

Possible way to run 2 BOINC installations on the same host is to use
VM for CPU-bound BOINC apps and host OS for CUDA.
I did such installation with Win/Linux BOINCs running together on single host.
But there will be VM overhead.... not sure it's worth to try.
But for host that does SETI and only SETI MB and has CUDA capable GPU it can be good solution.

i think if you install boinc 6.2 and 6.4 in two folders then run them both ...

Can't be done.

1. BOINC checks if another client is already running and if it is, the 2nd started will give a message about that and not run at all.

2. There's only one registry entry for BOINC, so if you install 6.4 first, then try to install 6.2 in another directory, this install overwrites the registry settings of 6.4, including where the Data directory is ... if you then start 6.4, it will run with all the preferences of 6.2, in the directory of 6.2

Possible way to run 2 BOINC installations on the same host is to use
VM for CPU-bound BOINC apps and host OS for CUDA.
I did such installation with Win/Linux BOINCs running together on single host.
But there will be VM overhead.... not sure it's worth to try.
But for host that does SETI and only SETI MB and has CUDA capable GPU it can be good solution.

I was going to suggest just the same thing. I'm currently running 2 VM's on my quad-core Q6600. The box runs Vista 32-bit, but the 2 VM's are running 64-bit Linux. While I'm running the same project (Primegrid challenge right now) on all 3 "boxes", I could just as easily run the GPU SETI client on the "host" box, and be running AP units in the VM's. Host client is set to only run 1 processor, and the other 3 are split between the VM's. Works fine here.

Note: the ability to run 64-bit VM's with VMware under a 32-bit OS is dependent on your CPU and BIOS. VMware has a utility on their website that allows you to check this... VMware player is free.

---

Mark

Yes I run both on this host all but four on each suspended it usually runs all setiBeta till complete then it runs Setimain till its complete. By this time I will have resumed some of the WUs again and it swaps over again. Manual control.

Michael

Maybe I should canablaize the other GPU card I have and put two in the one machine to see what it would do in that case.

If you want to head on over to beta I've been doing just that...
Dual GPU Testing

~Bob

---

Do you Good Search for Seti@Home? http://www.goodsearch.com/?charityid=888957
Or Good Shop? http://www.goodshop.com/?charityid=888957

Now im running stock, but still i cant see any cuda application in task manager... :-S

Its running app 6.03 atm, while it should be running 6.05 or?

Boinc reports CUDA device found.
So how come my GPU isnt crunching?

I had to detach and reattach to get Cuda WUs

---

I had to detach and reattach to get Cuda WUs

I have done a lot more than that ... :)

It is looking more and more to me that the problem is that the project is returning the same error message regardless of the reason that the project is not issuing work. Which is almost worse than not issuing an error message at all.

We are like the pidgeon that develops elaborate procedures to get the grain pellet when it is rewarded randomly. I also posted my dong Pixie's behavior on the GPU Grid's board ...

---

My question adressed for multicore system owners who run stock app.
Can you run simultaneously stock app on CPU _and_ stock CUDA MB app on GPU ?
(Not AP CPU + MB CUDA, but exactly CPU MB + GPU MB )

I haven't tried this and it probably won't work, but as they are different app versions you might be able to have in your app_info a version for seti_enhanced 603 (using cpu) and another one for 605 (using cuda). I suspect they will all come down and be tagged as 605's but you never know.

I'm testing that <ncpus>5</ncpus> in cc_config, that you suggested in another message thread, at the moment. I might give the 603/605 a try later on tonight unless someone beats me to it.

---

BOINC blog

My question adressed for multicore system owners who run stock app.
Can you run simultaneously stock app on CPU _and_ stock CUDA MB app on GPU ?
(Not AP CPU + MB CUDA, but exactly CPU MB + GPU MB )

I haven't tried this and it probably won't work, but as they are different app versions you might be able to have in your app_info a version for seti_enhanced 603 (using cpu) and another one for 605 (using cuda). I suspect they will all come down and be tagged as 605's but you never know.

...
I might give the 603/605 a try later on tonight unless someone beats me to it.

When using an app_info.xml, the Scheduler checks that a suitable version of the app is available before sending work, but doesn't tell the core client which to use. The core client does always choose the highest version from the app_info.xml, though.

The idea should work if you're ambitious enough. First, the optimized 603 executable will need to be mentioned with a <file_ref> in the 605 section to make sure the core client won't delete it. Crossreferencing the other direction should only be needed if you ever got to the situation where all 605 work had been completed, uploaded, and reported (but I'd include it just for the slight chance that might happen). Then after downloading a batch of work you could shut down BOINC and edit the client_state.xml so that a suitable part of the new work is assigned to version 603. Also check for any coprocessor/CUDA information attached to those results and edit it out, though I don't think there will be any. When you start BOINC again it ought to start both 603 and 605 work.

==============================================

The following email from the boinc_alpha mailing list gives an indication how long it may be before there's code in BOINC itself to handle the situation:

From: David Anderson <davea@ssl.berkeley.edu>
To: Eric J Korpela <korpela@ssl.berkeley.edu>
Subject: Re: [boinc_alpha] Seti@home running under CUDA ( way to cool )
Date sent: Fri, 19 Dec 2008 10:27:50 -0800
Copies to: boinc_alpha@ssl.berkeley.edu

Eric J Korpela wrote:
> I'm also getting requests from people who want to be able to run the CPU
> and GPU based applications simultaneously.

This is a month or two off.
Currently the scheduler figures out which app version is fastest
on your computer, and sends you jobs only for that version.
i.e. if you have a GPU you get only GPU jobs.

==============================================

                                                                Joe