[RC5] gpu client
kgmccoy at idk-inc.com
Wed Feb 18 11:17:40 EST 2009
> I'm not sure, but I believe the problem is not so much in the client
> being "not nice" in terms of scheduling with the GPU, but that the
> driver has to reload lots of stuff to the card's memory when it
> switches between graphics and the CUDA work.
It might be slow because this machine is running dual monitors at 1680 x
1050 x 32bit with a 60 Hz refresh. Whats that work out to? 3360x1050?
Eek! This is my software development machine (no games, some CAD) and I
like to have a lot of windows open at a time.
> At least to me it doesn't look like there is any difference between
> what is being drawn in 2D, the refresh rate just sucks when the CUDA
> Dnetc is running. The Vista Aero UI is another issue though.
I'm running XP - too many issues on Vista that I am not willing to deal
with. I'll let you early adopters bash MS bugs for me :-)
> Btw, what kind of keyrate do you get with the card and which core is
> the best for it?
It peaks out at ~80 MKeys/S
<http://idk.serveftp.net/cowboy/u1/host.html>. Click on the link and
look down my herd stats table until you see "Samovar" (get it? :-) )
Samovar is pretty much my only RC5-72 machine right now, since I
concentrate almost all my other cows on OGR.
The RC5-72 /average/ is a bit less, since I have to pause the thing to
get any work accomplished.
I am using the CUDA 2-pipe 64-thd core. It worked out best in the long
benchmark test I ran. Surprisingly, the short benchmark picked another
core that wasn't nearly as fast. I think the CUDA client developers need
to tweak the default-core picker a bit.
Kevin G. McCoy
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the rc5