CUDA/video card performance?

[QuietIce]

I'm not much on video card performance since most of my games are simulations, which require a lot of video memory but not a lot of speed, so I'm kinda' lost here trying to figure all this out.

I'm curious about the performance of a given video card for CUDA use. I don't use higher-end video cards for games, etc, so this would be a SETI only purchase to push the RAC on some of my older systems. I found a couple of graphs on the message boards showing 9800 and 8800 performance but, to tell the truth, I didn't see a lot of difference in the 9800 cards and half the time the 8800 seemed to be on par with them. From that information I assume video performance (ala 3DMark) isn't a big indicator of SETI performance so I'm trying to figure out what the most cost effective card would be.

I'm also wondering if PCIe (as opposed to PCIe 2.0) will radically effect SETI performance. Given the low CPU use for CUDA I assume the CPU passes off a chuck of data to the GPU, which crunches it then hands it back. If the data exchange between the CPU and GPU is low then the PCIe bandwidth and bus type, except for compatibility issues, may not make a huge difference???

A little insight please?


Thoughts, observations, experiences - anything goes ...!

 

[Duner]

http://www.overclock.net/overclock-net-folding-home-team/439190-great-news-pci-e-x1-folding.htmll

That threat shows folding on the GPU is not bandwidth limited by even a PCI-E x 1 slot, albeit on a rather slow card. I don't think PCI-E 1 vs 2.0 will make much difference at all on a x16 slot.

Seems like the 8800GT/9800GT G92's are the sweet spot for price/performance right now. When I bought mine, it was just for SETI. I don't even game on the PC anymore.

Also, Most New 8800GT are THE SAME as the 9800GT's. SOME newer 9800GT's are 55nm process, but that's hit and miss. Same clocks (unless factory OC'd), same memory, same shader count, same bus width.

 

[JDRacer47]

let me tell you

I am not a super techno guy but here is a little fact. I am running SETI on an AMD FX60 processor on a 3rd PC (1 core 100% of the time so the wife can use the other and not complain), and recently got my CUDA going on an 8800 GTX that I bought the first week they came out. My RAC has jumped 1000 points at least, and for an apples to apples comparison, a 14.4 credit WU takes up to 67 minutes for that FX60 core to crank out, my CUDA cranks it out a comparable WU in 43 SECONDS. I was like Holy Crap, I realized it was pretty beefy but had no idea. I don't even use my primary PC's processor anymore (e6600 @ 3.5 GHZ) because I get way more out of CUDA.

 

[Duner]

FYI, the 43 seconds is CPU time for the CUDA workunit. BOINC does not count GPU time for workunits. The only way to tell how long a CUDA unit takes is to use a stopwatch.

 

[JDRacer47]

FYI, the 43 seconds is CPU time for the CUDA workunit. BOINC does not count GPU time for workunits. The only way to tell how long a CUDA unit takes is to use a stopwatch.

Thanks, I am very new to this. My RAC jumped significantly anyway. My bad

 

[Rickdstik]

So, what are average WU times for say a x800 card. I kept wondering why RAC wasnt really (like much more than it is already) going up with the CUDA client. Im almost positive the 4870 will get a twin brother when the i7 box is completed, so I may stuff a 9800 in Skeletor.

 

[eaglescouter]

So, what are average WU times for say a x800 card. I kept wondering why RAC wasnt really (like much more than it is already) going up with the CUDA client. Im almost positive the 4870 will get a twin brother when the i7 box is completed, so I may stuff a 9800 in Skeletor.

http://setiweb.ssl.berkeley.edu/forum_thread.php?id=51712

 

[QuietIce]

So, what are average WU times for say a x800 card. I kept wondering why RAC wasnt really (like much more than it is already) going up with the CUDA client. Im almost positive the 4870 will get a twin brother when the i7 box is completed, so I may stuff a 9800 in Skeletor.

CUDA will run on ATI now ...?

 

[eaglescouter]

No, the CUDA software only works on nVidia cards.

ATI is rumored to be developing their own software, but there is no proof that I know of.

 

[Rickdstik]

by x800 i meant 8800/9800 ... though i'm drooling at the thought of an ati version, especially with the 4870 coming down in price so much of late.

 

[QuietIce]

*light-bulb* That makes more sense. Couldn't really believe you didn't know but ...

 

[Rickdstik]

I'm reading through that thread, and just need some clarification on that. When the poster referred to a 2.7 WU, those are the shorty units that grant 14.xx credit or so, right? Seems like around 7 minutes on the GTs, which is about half what my q6600 is doing them right now. So 4th core (dedicated to GPU only) would be generating about 2700 RAC alone, which would put total RAC at (rough estimate) around 7k. Not bad at all, I think I know Skeleltor's fate once the i7 rig (yet to be named, but definitely will not be tied to its case) is finished. Off to go cheap 9800 hunting.