GPU BOINC Basics
11 Dec 2010 14:33:44 UTC
Topic 195504
(moderation:
In putting multiple GPU cards into a BOINC machine are their any compatibility issues to consider?
Assuming one is using the same flavor (NVIDIA or ATI) for all of the cards in a particular machine.
For example, do the GPU models/generations have to be the same like with SLI/Crossfire? Or, can I put an early cuda NVIDIA card in with a NVIDIA Fermi, for example?
Thanks,
Steve
GPU BOINC Basics
)
It is best to put in the same kind but yes some people do get away with different kinds, some have even gotten Nvidia and ATI cards to both work in the same machine. It would be best if you went to a site like Collatz or Dnetc, which are mostly gpu based crunching projects but do cpu crunching on the side. The info about the mechanics of it would be easier to find over there. Before you go out and get a Fermi card though, make sure it will work, apparently Fermi cards do not crunch with all projects without some modifications to the software. Currently only Nvidia cards will work here on Einstein, so do not get an ATI card if that is going to be your focus. Although there is new software in the works here that is in the testing phase now, no time table for its release has been announced. Or whether ATI cards will work here in the future.
One of the things noticed with different models of cards in the same machine is Boinc can have troubles with the work units, one card is faster than the other and Boinc tends to wait on the slower card to finish before the faster card gets work again. In the end it is definitely easier to only put one card per machine, but of course that don't always work for everyone.
Mikey, Thanks for the
)
Mikey,
Thanks for the response. There certainly seems as much art as science in getting the GPUs to run with the various projects.
When I first returned to BOINC recently, my first machine had 2 older Nvidia cards (an 8400 and 8800) which I was able to get to work together. Emboldend by that I got a GTX460 which I swapped into the first machine and then got a second one for an additional machine. However, when I tried to add in the older cards for extract cuda power, I couldn't even get a screen. Fortunately, once I removed the older cards, no harm done, but it was what caused my question.
Regards,
Steve
RE: Mikey, Thanks for the
)
You are correct, these are very early days of gpu crunching and people are trying ALL KINDS of different combinations of cards. Anyone with more than one gpu card in their machine is cutting edge these days, although those with 3 or more are REALLY cutting edge!! Good luck with your setup.
I guess that puts me on the
)
I guess that puts me on the Hemorrhaging edge as I have been running as many as 5 GPU's. 2 Geforce GT 240's, 2 EVGA GT 240's and a Zotac GT 220 at the same time on a MSI 890FXA-GD70 m/b, with a 2.8GHz AMD Phenom II X4 820, 2G ram and a Corsair GS700 power supply for those that are interested. This runs at 4.4TFLOPS+ per kilowatt for those with a power budget. Watch out for "Windows Updates". The last update updated my NVIDIA drivers and I have had nothing but grief since.
Jim
I've just installed a GTX 460
)
I've just installed a GTX 460 graphics card in a new computer I built yesterday. Rather than using the provided driver CD, I went directly to the nVidia driver website and used the most recent non-beta driver for this configuration (Windows 7, 64-bit), version 280.26.
Over at SETI, I ran the excellent Lunatics unified installer, ticking the box for CUDA work, and almost immediately BOINC began requesting and receiving CUDA work. Limited by transfer reliability, it nevertheless has reported completion on over a dozen results in the past couple of hours, and one has validated already.
Here on Einstein, I notice in the message log of boinc that my host has requested and received new CPU work more than once, but does not seem to have requested, and certainly has not received, any GPU work. I am not running anonymous platform. I currently have all applications ticked in the "run only" section of Einstein preferences for the location (venue) used by the host. The "to send" column on the server status page shows non-zero work available. Other hosts have received cuda BRP work in this period.
Here is the SETI listing of this host.
Here is the Einstein listing of this host.
Any advice (including "just wait") would be appreciated.
RE: Any advice (including
)
Just waiting would not have worked. I made the most basic of errors--not setting the "use NVIDIA GPU" preference item to Yes in Einstein preferences for the location of this host.
On doing so, and forcing an update so that my host would read preferences, my host made an NVIDIA GPU work request after the one minute mandatory delay, and was promptly awarded 24 tasks of Binary Radio Pulsar Search work for the GPU, which are downloading now.
Sometimes it is the most obvious thing--my face is red.