jshay

Why are ATI cards bad for folding? Just wondering.

__________________
What's next?
1) Fermi
2) 160GB Intel X25-M SSD
3) New Mouse Pad

sdla4ever

CUDA is more developed then ATI's version of GPU acceleration

__________________

skitzogreg

All answers here.

http://theovalich.wordpress.com/2008...uture-reveale/


EDIT: While this is from 2008, the same theory applies with the memory. The 4800+ cards never really fixed this issue.

kz26

The F@H GPU client is not well optimized for ATI cards. If it was, you'd be seeing some mean PPD, especially out of the new 5800 series.

__________________

Afrodisiac

Inefficient client. People say it's because ATI doesn't put enough work into the client and optimization. I don't know about that though. There was something about a form of cache that allows the results to be valid on video cards.

ATI actually gets more PPD but a lot less of it is valid and actually registers because of the cache thing, I think. I'll dig up the info if I can. That's why the HD5* will supposedly destroy with the GPU3 client. They fixed that cache thing.

jshay

Heck yeah if that GPU3 client does rock then I'd be happy to fold 24/7. Right now I'm getting junk really :|

__________________
What's next?
1) Fermi
2) 160GB Intel X25-M SSD
3) New Mouse Pad

jpz

From what I have gathered, the issue lies with Brooke+ and its inability to access protected memory. Since protected memory can be accessed with CUDA for the nVidia client, all the data generated on nVidia cards gets stored in protected memory where it can later be retrieved if the client encounters an error. The nVidia client simply resumes work from these 'restore points', while the ATI client has to start all over from scratch.

There are also a number of software optimizations that have been applied to the nVidia client which could but have not been applied to the ATI client. The biggest one, as I recall, had something to do with 'scatter'. By using this 'scatter' feature, the nVidia client was able to cut the necessary mathematical operations in half. ATI cards have the same feature and Brooke+ supports it, but the F@H developers have not applied this optimization to the ATI client for some reason.

The GPU3 client is being written from the ground up using openCL and openMM, which means the GPU client will finally be able to access protected memory on ATI cards. Also, it is possible that we may have one unified GPU client for both ATI and nVidia cards, which would mean both nVidia and ATI owners could benefit from the same optimizations in the future.

__________________

jshay

When does the GPU3 come out? Any ideas?

__________________
What's next?
1) Fermi
2) 160GB Intel X25-M SSD
3) New Mouse Pad

jpz

Q1 2010 at the very earliest, but most likely not until long after. The F@H team won't give us a rough timeframe because there is still a lot that could go wrong, potentially causing massive delays. Last I heard ATI hadn't even gotten their openCL driver approved yet. GPU3 is coming, but not in the immediate future. You'll have to keep your eyes open for updates.

__________________