Author |
Message |
MJHProject administrator Project developer Project scientist Send message
Joined: 12 Nov 07 Posts: 696 Credit: 27,266,655 RAC: 0 Level
Scientific publications
|
Hi,
Several of you have asked exactly which graphics cards are supported by GPUGRID. All Geforce 8, 9 and 200 series GPUs are capable of running CUDA code. In practice, however, only those with 96 or more stream processors are fast enough to participate usefully. Currently, this means a 8800, 9800 or GTX260/280 card.
You may look up the configuration of a particular card on the following wikipedia pages:
Currently we use 8800GT and 9800GX2 cards for our development.
Note that because the GPUGRID code works the graphics card very hard, we do not recommend using an overclocked card.
Edit 25 Aug 2008: We are not currently able to support the 320MB and 640MB 8800GTS/X cards.
MJH
|
|
|
|
Hi,
Several of you have asked exactly which graphics cards are supported by GPUGRID. All Geforce 8, 9 and 200 series GPUs are capable of running CUDA code. In practice, however, only those with 96 or more stream processors are fast enough to participate usefully. In practice, this means a 8800, 9800 or GTX260/280 card.
You may look up the configuration of a particular card on the following wikipedia pages:
Currently we use 8800GT and 9800GX2 cards for our development.
Note that because the GPUGRID code works the graphics card very hard, we do not recommend using an overclocked card.
MJH
Now you should explain me why those VGAs with less than 96 stream processors are unable to "partecipate usefully". Weaker cards will just take more time to process datas, isn't it?
In our team, Boinc.Italy, we have many users with 8400 and 8600 cards that WANT to partecipate but CAN'T and this, in my humble opinion, is like WASTING some good computing power. Do you know how many Pentium 2, Pentium 3, Athlon XP and Duron are still out there running BOINC? Are they not "partecipating usefully"? |
|
|
Saenger Send message
Joined: 20 Jul 08 Posts: 134 Credit: 23,657,183 RAC: 0 Level
Scientific publications
|
In our team, Boinc.Italy, we have many users with 8400 and 8600 cards that WANT to partecipate but CAN'T and this, in my humble opinion, is like WASTING some good computing power. Do you know how many Pentium 2, Pentium 3, Athlon XP and Duron are still out there running BOINC? Are they not "partecipating usefully"?
I've got an 8600GT running this project, it crunched the first WU without problems. Now is a sample of 1 not really sufficient for general remarks, but I'd say: let them give it a try!
____________
Gruesse vom Saenger
For questions about Boinc look in the BOINC-Wiki |
|
|
|
A longer deadline would be nice.
Certainly for when you make a windows version available, as you are going to have a lot bigger user base, and probably by definition a lot more complaints about the deadline being too short.
____________
Down with the Kredit Kops!!! |
|
|
GDFVolunteer moderator Project administrator Project developer Project tester Volunteer developer Volunteer tester Project scientist Send message
Joined: 14 Mar 07 Posts: 1957 Credit: 629,356 RAC: 0 Level
Scientific publications
|
A longer deadline would be nice.
Certainly for when you make a windows version available, as you are going to have a lot bigger user base, and probably by definition a lot more complaints about the deadline being too short.
The type of application that we provide require a lot of computer time, this is why it is PREFERABLE to join if you have at least 8800GT.
You can still try to return workunits with 8600 type cards, but it will take you much longer. 8400 are 20 times slower than top cards and cannot return wu within the deadline.
They can still connect as we do not have any mean to exclude them. This is why we specified better what are the computational resources recommended.
GDF
|
|
|
Saenger Send message
Joined: 20 Jul 08 Posts: 134 Credit: 23,657,183 RAC: 0 Level
Scientific publications
|
The type of application that we provide require a lot of computer time, this is why it is PREFERABLE to join if you have at least 8800GT.
You can still try to return workunits with 8600 type cards, but it will take you much longer.
If I get the numbers right from this thread, and am right to assume that CPU doesn't matter that much, the difference is about 4 times. The WUs take about 12h for UBT - NaRyan on a 8800 and 50h on my 8600.
____________
Gruesse vom Saenger
For questions about Boinc look in the BOINC-Wiki |
|
|
MJHProject administrator Project developer Project scientist Send message
Joined: 12 Nov 07 Posts: 696 Credit: 27,266,655 RAC: 0 Level
Scientific publications
|
[A]m right to assume that CPU doesn't matter that much, the difference is about 4 times. The WUs take about 12h for UBT - NaRyan on a 8800 and 50h on my 8600.
That's correct. In general, the performance of the GPUGRID application will scale linearly with the GLFOPS rating of the GPU (refer the Wikipedia articles above for these values).
MJH |
|
|
|
A longer deadline would be nice.
Certainly for when you make a windows version available, as you are going to have a lot bigger user base, and probably by definition a lot more complaints about the deadline being too short.
The type of application that we provide require a lot of computer time, this is why it is PREFERABLE to join if you have at least 8800GT.
You can still try to return workunits with 8600 type cards, but it will take you much longer. 8400 are 20 times slower than top cards and cannot return wu within the deadline.
They can still connect as we do not have any mean to exclude them. This is why we specified better what are the computational resources recommended.
GDF
Exclude? Are you serious?
People want to help and you want to EXCLUDE them?
What's better? More power in a longer time or less power in a shorter one? |
|
|
|
Maybe you should read GDF's post again?! ;)
They can still connect as we do not have any mean to exclude them
Also we are still in the early testing stage and GDF and his team will for sure choose what's best for their project...
____________
pixelicious.at - my little photoblog |
|
|
Saenger Send message
Joined: 20 Jul 08 Posts: 134 Credit: 23,657,183 RAC: 0 Level
Scientific publications
|
Exclude? Are you serious?
People want to help and you want to EXCLUDE them?
What's better? More power in a longer time or less power in a shorter one?
My old Commodore C128 is excluded from crunching for BOINC, I doubt whether my first real PC, a 286er, would not be excluded as well.
Imho people have to live with the simple fact that computer hardware is outdated quite fast. It may still work as a browsing machine, but will not do anything useful for science (besides wasting energy). It's a bit like demanding that the Formula 1 racing rules should be inclusive for the Käfer/Beetle.
____________
Gruesse vom Saenger
For questions about Boinc look in the BOINC-Wiki |
|
|
|
Maybe you should read GDF's post again?! ;)
They can still connect as we do not have any mean to exclude them
His knowledge of the fact means that they actually thought about (or even tried to) excluding them.
Anyway, I understand that some cards may be slow. A 8400 will take around 250 hours for short WUs and 650 hours for long ones. That means 11 days and 30 respectively. Maybe a bit too much, considering that it's not sure those PCs will run 24/7.
But a 8600GS or GT, with just 32 stream processors, will complete a WU in 50 (short) and 130 (long) hours. A couple days and less than a week (24/7). Let's say they'll run 8/5 (work time) and you'll get a WU completed in slightly more than a week and slightly more than 2 weeks respectively. Is that unacceptable for the project?
That's what I don't understand. Why is there the need for such a short deadline? Is it just for quick developement purposes or will it stay the same when the project will exit the alpha and beta status? |
|
|
FilipeSend message
Joined: 28 Apr 08 Posts: 3 Credit: 1,994,582 RAC: 0 Level
Scientific publications
|
Hi all,
i'm using a slightly over clocked 9600GT (XFX XXX version)on Ubuntu Hardy (8.04), cuda drivers 177.13, Pentium D processor. It takes about 18 hours to process, in line with what Zeuz said in an earlier post. The card has 64 streaming processors.
However my first two runs had computational errors (at about 52% and 90% completion ..if i recall correctly),the third will finish in 2-3 hours..... one way or another.
I'll post an update after completion of the 3rd unit, with some other issues i came across.
Cheers,
Fil.
p.s the first error was my fault as i accidentally hit the sleep mode, and obviously the gpu didnt like it as it didnt pick up from where it left off.
|
|
|
GDFVolunteer moderator Project administrator Project developer Project tester Volunteer developer Volunteer tester Project scientist Send message
Joined: 14 Mar 07 Posts: 1957 Credit: 629,356 RAC: 0 Level
Scientific publications
|
Hi all,
i'm using a slightly over clocked 9600GT (XFX XXX version)on Ubuntu Hardy (8.04), cuda drivers 177.13, Pentium D processor. It takes about 18 hours to process, in line with what Zeuz said in an earlier post. The card has 64 streaming processors.
However my first two runs had computational errors (at about 52% and 90% completion ..if i recall correctly),the third will finish in 2-3 hours..... one way or another.
I'll post an update after completion of the 3rd unit, with some other issues i came across.
Cheers,
Fil.
p.s the first error was my fault as i accidentally hit the sleep mode, and obviously the gpu didnt like it as it didnt pick up from where it left off.
I have been seeing your errors. The compute errors seem to be correlated to the card/driver of each machine as some users experience none, some others a lot.
We still don't know if there is a workaround it, so keep us informed.
We probably have found a solution for the libcudart.so error which should be fixed in the next release of the BOINC client.
gdf
|
|
|
|
hi, i need to choose a laptop and i want to run gpugrid on it
do you know the number of SP of the quadro3600M and the 1600M ? |
|
|
|
Maybe you should read GDF's post again?! ;)
They can still connect as we do not have any mean to exclude them
His knowledge of the fact means that they actually thought about (or even tried to) excluding them.
Anyway, I understand that some cards may be slow. A 8400 will take around 250 hours for short WUs and 650 hours for long ones. That means 11 days and 30 respectively. Maybe a bit too much, considering that it's not sure those PCs will run 24/7.
But a 8600GS or GT, with just 32 stream processors, will complete a WU in 50 (short) and 130 (long) hours. A couple days and less than a week (24/7). Let's say they'll run 8/5 (work time) and you'll get a WU completed in slightly more than a week and slightly more than 2 weeks respectively. Is that unacceptable for the project?
That's what I don't understand. Why is there the need for such a short deadline? Is it just for quick developement purposes or will it stay the same when the project will exit the alpha and beta status?
I'm quoting myself because I'm still waiting for an answer.
I bought a 8800GT and I'm running this project 24/7, do I deserve an answer now?
|
|
|
GDFVolunteer moderator Project administrator Project developer Project tester Volunteer developer Volunteer tester Project scientist Send message
Joined: 14 Mar 07 Posts: 1957 Credit: 629,356 RAC: 0 Level
Scientific publications
|
That's what I don't understand. Why is there the need for such a short deadline? Is it just for quick developement purposes or will it stay the same when the project will exit the alpha and beta status?
The few days deadline will stay as the aim of this project is to create by distributed computing the largest supercomputer for applications of biomolecular simulations. Our calculations are so heavy that they will require weeks on slow machines. Such an high time to return the workunits is not ideal for distributed computing as the resources are volatile. One day of calculations is probably the maximum. Also, it is crucial to reduce the latency between submission and return of the results to be analyzed. So, we do not need the biggest possible pool of machines, but the fastest pool of machines.
Hope it helps.
gdf |
|
|
|
Hi,
I'm looking to buy a new card and I'd like to know if i should buy a
8800GT 256 or 512MB?
I found a 8800GT 256 at $60
What is the best for GPUGRID?
Thank you
Eric |
|
|
GDFVolunteer moderator Project administrator Project developer Project tester Volunteer developer Volunteer tester Project scientist Send message
Joined: 14 Mar 07 Posts: 1957 Credit: 629,356 RAC: 0 Level
Scientific publications
|
Hi,
I'm looking to buy a new card and I'd like to know if i should buy a
8800GT 256 or 512MB?
I found a 8800GT 256 at $60
What is the best for GPUGRID?
Thank you
Eric
The 256MB one should be fine for current version of the application, but we cannot guarantee for the future.
gdf |
|
|
|
Two questions:
Does GPUGRID support the Telsa C870?
Does GPUGRID support using multiple video card on the same host? Specifically, I'm looking at 2 of the Geforce GTX 280. OK, I found this thread, so if I have four CPUs and four video cards, then I can run 4 GPUGRID tasks simultaneously?
I'm running Fedora 8 on x86_64. |
|
|
GDFVolunteer moderator Project administrator Project developer Project tester Volunteer developer Volunteer tester Project scientist Send message
Joined: 14 Mar 07 Posts: 1957 Credit: 629,356 RAC: 0 Level
Scientific publications
|
Two questions:
Does GPUGRID support the Telsa C870?
Does GPUGRID support using multiple video card on the same host? Specifically, I'm looking at 2 of the Geforce GTX 280. OK, I found this thread, so if I have four CPUs and four video cards, then I can run 4 GPUGRID tasks simultaneously?
I'm running Fedora 8 on x86_64.
Hi,
if you are happy to test, I can upload a new Linux application which should automatically support multiple GPUs. If you have 4 cpus and 4 GPUs cores, you can have 4 GPU wu running. Four G280 should deliver something of the order of 20,000 credits per day. However, some users had problems with the driver of the 200 series.
Yes, we do support tesla units too, but they are not faster than graphics cards.
gdf |
|
|
|
Hi,
if you are happy to test, I can upload a new Linux application which should automatically support multiple GPUs. If you have 4 cpus and 4 GPUs cores, you can have 4 GPU wu running. Four G280 should deliver something of the order of 20,000 credits per day. However, some users had problems with the driver of the 200 series.
Yes, we do support tesla units too, but they are not faster than graphics cards.
gdf
Thanks for the reply. I'll run my test machine for a little while until I can get $2,000 to buy my new super-computer. :)
|
|
|
|
Hi, think about this:
When the project is really up and running, with proper research being done, woudn´t it be nice to have different experiments running on different types of machines?
I undestand that you want the fastest machines (So far I myself am included on those, with a 96sp 9600GSO only for modeling, and a 6800gs for the video rendering, probably two 260 to come next year) so you have the fastest grid, but as the BOINC idea is to run tasks on background while people are using their computers the energy used for the projects is only a 15% to 40% excess when compared to that of someone just navigating and it is not really a waste of energy (it would be if you left a 400watts machine running a 8400 just for the project though).
So I recommend (and if possible would like to help) producing a way of knowing the card the user has and dividing the researches in many parts, some of them with longer deadlines running strictly on weaker machines and some of them running only on the best ones, so you have quick results on the research you want to and still can run other thing that are "less important" (in a good way, things that would go to the end of the project queue) that can wait for a while.
I believe it is important to use all the avaiable idle cards because it is computing power and usefull.... as an example, I have a notebook with a 8400m gs, I will not try running GPUGRID now on it because it has only 16sp, but if there were lighter tasks with longer deadlines I sure would... why? Because I use my note all day long and only with 2d app, the GRID would be running on this 8400gs that uses only about 30watts of power!! And because it already uses about 15-20 watts all the time, it would be a very energy efficient way of running the grid....
Last thing, ok, my 8400gs would do the job 50 times slower than a GTX280... but go ask dell, apple and HP how many notes and desktops with "weak less than 96sp graphics card" they sell a week and you might discover that the amount of computing power wasted there is much greater than the couple hundred of GTX´s that are sould o the same time.... and all those "weak" computers will be run at some 8 to 10 hours a day!!
That is it, I will continue to support you guys (I´m travelling for 2 weeks now and stopped receivig tasks from you and from WCG, but will retake on january), but I ask you not to diminish the weaker hardware as it will be around for quite a while... if BOINC did not support 486´s at the beggining or pentium 3´s or athlon XP´s today it would not grow that much, as people would fell regected by those they wanna help and when they have a hadware that you want they will not help....
Remember that not everyone who are on grids are geeks that dispose their money on the top of the top just to run for points, most of them really wanna help (not that the geek doesn´t, but he has the competition excitement) and I think they can!!
att
Leandro Biondo
____________
|
|
|
|
I am very amused about the discussion about graphics card and why not have longer deadline. My opinion is clear. I do not buy graphic cards especially for crunching, I buy which I need for my work or for my games. If they meet the requirements for a project it is fine, if not I do not have a problem.
The team of GPUGRID has defined which card is good for the project and which is too slow. There are many other projects you can help with you CPU-resources if you don't have a card which is requested.
I have 2x8800GTX/768 "on stock" in one of my PC's, it does not meet the minimum requirements. I am running a notebook with Nvidia onboard card, it does not meet the minimum requirements. I don't have a problem with it.
I have two PC's for gaming, when the PC is running and I am not in a game then i activate GPUGRID. And of course other projects running on the CPU. When I play games the work is suspended. I think most of the user don't let their PC's run only for projects, so think of the long time the WU's need if not running 24/7 only for projects.
So please be fair to all the project teams, the deadlines they give us and do not ignore their minimum requirements! I think they know why they did this. If we can help them with our resources, let's do it!
The last to say is "merry christmas and a happy new year"!
____________
|
|
|
|
Leandro,
I'm sure your feedback is welcome.. and I'm also sure it's somewhere on the developers to do list to become more flexible in the type of work they do. I just think it's not very high up on the list. If I were them I'd get the stuff neccessary for high performance crunching done first and care about the rest later.
Regards,
MrS
____________
Scanning for our furry friends since Jan 2002 |
|
|