Seti@Home optimized science apps and information
Optimized Seti@Home apps => Windows => Topic started by: Frizz on 10 Feb 2010, 03:26:27 am
-
Hi there,
I would like to set up a dedicated host for BOINC. I was thinking about a ATI 4870 or 5770 card.
The 5770 is more energy efficient - but lacks DP!
My question: How does the SP version of "HYBRID ATI/CPU AstroPulse" compare to the DP version?
cheers,
Frizz
-
DP is slightly faster.
But both if you want SETI cruncher, better buy nVidia GPU now.
HD5xxx will be supported more fully than now but later...
And BTW, HD4xxx lack of shared memory support - no future for those GPU it seems at least on SETI project :(
-
Thanks for the reply.
I thought I go for an ATI card this time because it looks like ATI is "doing more" than nVidia when it comes to Open CL and Open GL. So maybe the (near) GPU future belongs to ATI?
What do you think?
-
I think that ATI still has worse support for their own hardware than NV. Regarding OpenCL too.
First of all HD4xxx don't supported by OpenCL properly and never will be, this haedware doesn't map onto OpenCL. All NV cards based on CUDA map perfectly.
Also, current OpenCL from ATI has some bugs re GPU (perhaps NV implementation has some bugs too, but ATI bugs showstopper for SETI currently).
That is, RAW-performance not all that needed actually. W/o good software what one can do with that performance? ... :-\
-
I think that ATI still has worse support for their own hardware than NV.
I guess one can agree here.
But have you seen ATI's OpenCL SDK 2.01 came out yesterday? Beside some bugfixes there is now a FFT sample application included. AFAIR you struggled a lot with the missing FFT routines for ATI. Maybe it's now worth a second look? I didn't bother with OpenCL so far (simply had other priorities), but I guess it will be the way to go in the future, even when only HD5000 GPUs will be properly supported. Furthermore, CUDA apps can be ported often in a quite straightforward way.
-
Yes, I already aware about new SDK and FFT sample, thanks :)
Will look at it. Currently I use Apple's OpenCL FFT sample that works on NV, on CPU but not on ATI GPU. Hope ATI's own sample will work on their hardware correctly ;D
Welcome here, btw :)
-
1) provided by AMD FFT sample just hardcoded size of 1024 FFT - not suitable.
2) variable-size FFT provided by Apple still produced incorrect results on GPU with SDK 2.01
-
I recently stumbled upon some speedy FFT code by Vasily Volkov:
http://forums.nvidia.com/index.php?showtopic=69801
It's CUDA.
Just thought I might post this here.
-
Cheers Frizz23. If you're using last release V12, you have a test (experimental) kernel which is a Volkov FFT size 256 with integrated powerspectrum. Cuda V2.3 DLLs made a significant FFT upgrade shortly after this release that did make the FFT less important, but still very interesting. I'm finding nVidia's clowning around with facebook announcements pretty irritating at this time, so please excuse me if I dance around threatening to trade in my nVidia registered developer credentials for a hamburger with cheese ...
Jason
-
I'm finding nVidia's clowning around with facebook announcements pretty irritating at this time, so please excuse me if I dance around threatening to trade in my nVidia registered developer credentials for a hamburger with cheese ...
Yeah me too. Looks like they have huge problems with Fermi. But instead of coming clear they clown around, rebrand their old crap (GT330), ...
I would love to switch to ATI - but their drivers are even worse than nVidias.