This is probably an oft-discussed topic but I can’t find it in the groups.
Initially I installed the nVidia driver from the CUDA web page, then I reinstalled a newer driver from the general nVidia page:
cudadriver_2.3_winxp_32_190.38_general.exe: processing times are great, with little variations from iteration to iteration.
195.62_desktop_winxp_32bit_english_whql.exe: processing times are poor, sometimes with wild variations from iteration to iteration.
Is this expected? That’s going to be a maintenance issue because video drivers are easily updated by anyone, but the relationship with GPU performance might not be obvious.
There are a few known performance bugs in 195.62 that have been fixed. If you have specific cases that are obviously slower (when using the same binary, cubin, etc), please send them our way.
I stripped-down a test program to the minimum needed to show the problem. I removed external dependencies, the GUI sucks but it shows the problem (on my computer anyway). Just click on ‘create buffers’, then click on ‘process now’ repeatedly.
In the GPU processing I basically call the same function (Invert4) 3 times. Processing time varies between 3 and 25+ ms. If instead I call the function only once, processing time goes down to 0.17 ms.
I’m running Windows XP (32 bits), I ran tests with a GTX 260 and GTX 285. Please let me know if you need more information or if I can be of any help. TestCUDA_3.zip (472 KB)