How does BOINC / Einstein handle multiple GPUs?
10 Jan 2012 10:52:44 UTC
Topic 196134
(moderation:
I'm running a GTX 560 and I'm not very pleased with its performance. I'm thinking about purchasing a pair of 550Ti's and running them together in a single board. This would slow down the PCIex bus to 8x.
Now, what do you think, what would be faster - one 560 or a pair of 550s? Will the cards work in unison or will I see two CUDA WUs running together? And the bottom line, is the PCIex drop to 8x going to significantly affect the computation speed?
Edit: I just saw that my PCIex bus is already running at 8x speed.
My best,
Andrej from Croatia.
How does BOINC / Einstein handle multiple GPUs?
)
It's the latter. Each GPU will run a task.
Gruß,
Gundolf
Computer sind nicht alles im Leben. (Kleiner Scherz)
PCIe bus speed has almost no
)
PCIe bus speed has almost no affect on BOINC,
The amount of data is so small compared to game`s.
PCI-E bus does have an effect
)
PCI-E bus does have an effect with Einstein CUDA apps. I have seen at least a 30% loss in performance in both Linux and Windows by moving a GPU from an x16 slot to an x8 slot.
Well, I installed a pair od
)
Well, I installed a pair od 550Ti's, and performance is satisfactory. Not stellar, though - it takes around 25 minutes to crunch a CUDA WU, but my PCIex bus is already bogged down with an ethernet card and a TV card so it's not like that the second video card is to blame for the slowdown.
Will I benefit by overclocking the PCIex bus to say, 130MHz? Data throughput could be increased, but so far my experience has taught me that the only thing I'll achieve with this is a generation of LOTS of heat and no performance improvement. What do you guys think about this?
RE: What do you guys think
)
The prime benefit from GPU usage is the massive parallelism on the card, but of course that requires thread setup first ( thus includes data -> card ) and then the return leg back to system RAM ( a lesser load ). As usual experimentation is the arbiter .....
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
RE: Will I benefit by
)
Ive tried to OC the PCIe Bus and I was not able to rise it more than 105Mhz... Above that speed all the PCIe cards (including the onboard lan) started to fail.
Im currently running it at 101Mhz but only cause Ive readed that an even frequency avoids some of the failsafe clockdowns of the GPUs... Not sure if thats true but since Ive set it at that speed my cards are downclocked less often...
Also, in my testings, OC the CPUs is what gives better results for all Einstein's tasks. OC the GPUs gives no significative improovements unless you push the clocks very, very high, which can cause invalid results that will defeat the improovement in performance...
RE: I'm running a GTX 560
)
I have just read that the Intel "Sandy Bridge motherboards can run two 16X PCIe vidio cards but only at 8X, the Sabertooth 990FX (AMD) boards can run the same but in full 16X" (February 2012 Maximum PC mag)
I have been planning to build a monster crunching machine and in the course of research, found this interesting statment.
I had not planned on going with an AMD proc/MB, but I guess I'll have to go back to the drawing board.
RE: I have just read that
)
If you wait a couple weeks you should have the option for a x79/3820 combo. The Intel 3820 has support for 40 PCI-E 3.0 lanes and will be priced a bit less than the 2600K at $285. This processor should allow for x16/x16/x8 configuration. An early review shows that the 3820 can OC to around 4.7 GHz with a combination of up to a 43x multiplier and 1.25 bclk multiplier.
RE: Hwiparam Have tou
)
Have tou tried to run an app_info.xml?
If you run only 1 wu at a time, Yours 560/550 are probably not being used at 100% load.
I also have a 550ti (Running a few hours a day).
Running 1 WU takes +/- 1 hour, and 60% load on the gpu.
Running 2 WU simultaneously takes +/- 1.5h and 95% load on the gpu.
It Gives me 100% more output for only 50% more time. Do the math.
Filipe
RE: Have tou tried to run
)
Where do I find this file? What should I edit in once I locate it?
Back when I ran only one 560, it took me 35 mins to crunch a single CUDA WU. Now that I'm running a pair of 550s, crunching of a single CUDA WU takes around 48 minutes. With a pair of video cards, it comes down that one CUDA WU on my computer takes 24 minutes to crunch. Is there some room for improvement here or should I be satisfied with what I've got? =)
My best,
A.