Seti@Home optimized science apps and information
Optimized Seti@Home apps => Windows => GPU crunching => Topic started by: glennaxl on 21 Mar 2010, 01:41:49 am
-
CUDA 3 is finally out (http://blogs.nvidia.com/ntersect/2010/03/gpgpu-developers-get-boost-from-new-cuda-toolkit-30.html).
Download Link (http://developer.nvidia.com/object/cuda_3_0_downloads.html).
It doesn't seem to have that much of an improvement - 1% increase.
Here's my test:
driver : 196.75
CPU: 100%
GPU: gtx 295 core 0
[attachment deleted by admin]
-
Hi , after I put in a 2nd nVIDIA card, The OS (WIN XP64) immediatly responded and downloaded the latest drivers and CUDA 3.0.14.
I tried CUDA 3.0 before with driver 196.68/75 with the 9800GTX+, which became/was unstable. And no speed gain. (Could have been BOINC 6.10.18, too.) Using 6.10.17 now.
After adding a GTS250 (also a G92 chip), I downgraded the driver to 190.38 and CUDA 2.3 Libraries. After inserting a cc_config.xml, with use_all-gpus, as option and some debug-flags. (coproc; xfers), put a 2nd monitor to 2nd card an extend the desktop, to it. It works OK, after removing the 2nd monitor.
-
On my GT 240 it doesnot work- computation error!
So what part of CUDA 3.0 you use? And how?
Last time ( with beta 3.0) it was like this
What one needs to do to use current CUDA MB builds:
use fully named cudart DLL (that is, not cudart.dll, but cudart32_30_8.dll )
and use renamed CUFFT DLL (that is, rename cufft32_30_8.dll to cufft.dll ).
Of course I renamed files accordingly new names
-
On my GT 240 it doesnot work- computation error!
So what part of CUDA 3.0 you use? And how?
Last time ( with beta 3.0) it was like this
What one needs to do to use current CUDA MB builds:
use fully named cudart DLL (that is, not cudart.dll, but cudart32_30_8.dll )
and use renamed CUFFT DLL (that is, rename cufft32_30_8.dll to cufft.dll ).
Of course I renamed files accordingly new names
Good question. I wasn't paying attention to the error that I got but instead I used both, the original name and the renamed one.
like this:
cudart32_30_14.dll
cudart.dll
cufft32_30_14.dll
cufft.dll
BTW, are you using CUDA 3.0 capable driver?
-
In that case I am 100% sure that you dont use new drivers at all. Of course I use 3.0 capable driver ( 197.13)
Remove old drivers cudart.dll and cufft.dll and you will get computation errors.
So what is left is try to use new dll with stock application.
update: it doesn't work. It looks like cuda 3.0 dll are not compatible with current cuda app
-
In that case I am 100% sure that you dont use new drivers at all. Of course I use 3.0 capable driver ( 197.13)
Remove old drivers cudart.dll and cufft.dll and you will get computation errors.
So what is left is try to use new dll with stock application.
update: it doesn't work. It looks like cuda 3.0 dll are not compatible with current cuda app
197.13 is a beta driver. I use 196.75 which they removed it from production.
To make it clear here's what I did:
cudart32_30_14.dll
cufft32_30_14.dll
copy of cudart32_30_14.dll -> renamed to cudart.dll
copy of cufft32_30_14.dll -> renamed to cufft.dll
-
If use your method it works :)
-
The same way it works for COLLATZ!
Filename here: "cudart_CUDA23.dll"
-
Something amiss with 32bit and 64bit downloads. The 32bit rt and cufft part in 64bit installer (http://developer.download.nvidia.com/compute/cuda/3_0/drivers/devdriver_3.0_winvista-win7_64_197.13_general.exe) has a diff file size.
64bit installer:
cudart32_30_14.dll = 273KB
cufft32_30_14.dll = 12403KB
32bit installer:
cudart32_30_14.dll = 280KB
cufft32_30_14.dll = 12188KB
Code guru's, can you comment on this?
-
The CUDA_V3.0_final have improvement/speed up compared to CUDA_V2.3?
Someone made a test if the newest nVIDIA_driver & CUDA_V3.0_final is faster than nVIDIA_driver_190.x (or 191.x) & CUDA_V2.3 ?
Thanks! :)