Forum > GPU crunching
Latest drivers (NVidia and ATI)
Fredericx51:
Not very lucky, getting mostly VLAR tasks, marked _vlar, so nothing to do for the FERMI's except GPUGrid and Einstein.
And still using the 258.96. drivers, giving a better response, no 'lag' or late response from the mouse or
keyboard.
Since a lot of Einstein tasks, which errorred out, after I made an mistake in the app_info.xml, were
resend, I can do them again.......
M_M:
I have upgraded to 260.63 drivers, including nvcuda.dll (CUDA 3.2.1), supposed to be much better optimized for fermi then existing CUDA 3.0.
From app_info.xml I see that Lunatics_x32f_win32_cuda30_preview.exe is probably using it own (CUDA 3.0) dll's, cudart32_30_14.dll, cufft32_30_14.dll, libfftw3f-3-1-1a_upx.dll. Is there somewhere a CUDA 3.2.1 version of such dll's? Shall we soon expect 0.38 unified installer and optimized application based on cuda 3.2.1?
Jason G:
DLL swapping is not advised as was practical in the past. Nvidia has made them 'strongly versioned' as of recently, so jiggering things to use libraries other than what the application was built with is an 'at your own risk' affair... and is not guaranteed to work or be stable.
Cuda DLLs are, as always, found in the respective Cuda SDK.
There are a lot of improvements in Cuda 3.2, and I intend to make builds with it, though the benefits are mostly feature additions rather than performance of existing code, and use of these new features is yet to be implemented here. We skipped Cuda 3.1 because of specific bugs, which apparently are resolved in 3.2. That will allow development to continue, where it had stalled with Cuda 3.1 issues, forcing a return to Cuda 3.0.
Since the features that Cuda 3.2 adds require further development to make use of, it'll be 'a while' before any new release uses it. Not a lot of point making Cuda 3.2 builds that are the same speed as teh Cuda 3.0 builds & don't use the new features ;) Using those will take time.
Jason
Zeus Fab3r:
--- Quote from: M_M on 09 Oct 2010, 01:41:35 pm ---I have upgraded to 260.63 drivers, including nvcuda.dll (CUDA 3.2.1), supposed to be much better optimized for fermi then existing CUDA 3.0.
--- End quote ---
Pa mogao si i mene da pitaš ;D Jason ti je sve lepo odgovorio, nema više kao nekad skineš dll-ove i baciš u folder. Ja još uvek ne mogu da provalim ima li ikakve vajde da sa GTX260/191.07/cuda2.3 prelazim na 2xx.xx drajvere i x32f. Brzina je verujem ista a jedino se manje javljaju -12 greške. Nego kad pre nabavi Fermi ?ż :o
pozz, ZF
[edit]sorry guys, it's my fellow countryman[/edit]
MarkJ:
--- Quote from: Jason G on 09 Oct 2010, 02:06:06 pm ---<snip>
Since the features that Cuda 3.2 adds require further development to make use of, it'll be 'a while' before any new release uses it. Not a lot of point making Cuda 3.2 builds that are the same speed as teh Cuda 3.0 builds & don't use the new features ;) Using those will take time.
Jason
--- End quote ---
Would a cuda 3.2 version without any additional code changes make sense for the Fermi guys? Given the supposed speed increases on Fermi perhaps it might be worthwhile.
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version