Problems with in consistent speed Is it drivers or x64 or ???

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, I wasnt exactly sure where to post this so I put it under general

I have 3 very similar computers.

  1. Abit ab9 pro mb w/ intel Q6600 4gb of 800mhz ram
    single gtx260 at pcie x16 (1.0)
    windows 2003 (r2) enterprise edition

  2. Asus P5Q deluxe w/ intel Q9300 2gb of 1066mhz ram
    single gtx280 at pcie x16 (2.0)
    vista x64 (sp1) enterprise edition (yes enterprise is really the edition)

  3. Asus P5Q deluxe w/ intel Q9300 2gb of 1066mhz ram
    dual gtx260 at pcie x16 (2.0)
    windows xp (sp3) professional edition.

ok I have double checked that they all have the newest update for drivers

If I run code against these video cards, I get very surprising results
The fastest by at least double is computer 1.
The next is computer 3 (even when running only 1 card)
Finally the slowest is computer 2.

All the computers pass rigorous stress testing (including on the video card)
The computers have plenty of power (all three are powered by pc power silencer 750)

For example if I try the fluidsgl example I get frame rates of @ 1024x1024grid

  1. 60fps
  2. 8fps
  3. 25fps
    An example of my code (includes both cuda and non cuda code) The code is compiled for 32bit and 64bit
  4. 1hr 37min 18s
    1 w/o cuda) 3days 13hrs 27min 58s just to give an idea of the code complexity
    2 @64bit) 5hr 9min 22s
    2 @32bit) 4hr 45min 19s
  5. 3hr 0min 47s
    I would post the code however it is confidential and rather lengthy.

I would like the results to be a little more consistent (even closer would be nice)

I would not consider my self an expert on cuda, but I do understand computers and programming, especially c++.

I would have believed that computer 3 would have been the fastest then computer 2 and finally computer 1…

Anyone have any suggestions on why this isnt the case???

At first I thought it was drivers, but windows xp and windows 2003 use the same drivers.
Are the Vista drivers (and/or maybe just vista and/or the x64 part) that much slower than windows 2003?

Has anyone else had similar problems?

2

No. I get nearly identical performance for CUDA kernels on linux32, linux64, winxp, and vista64.

You mention that you have the “latest” drivers on each machine. Which CUDA toolkit are you using? Certain drivers only support certain versions of the CUDA toolkit. Specifically, if you are using CUDA 2.0 you need driver version 177.84 (http://www.nvidia.com/object/cuda_get.html).

Outside of that, I can’t help you besides giving advice. Whole program performance isn’t very useful in troubleshooting: you need to apply problem solving strategy #1, find the root cause. Profile different portions of your code and see which take longer on the various architectures.

You use fluidsGL for an independent test, but I don’t think it is very good for testing the speed of CUDA calculations. Since it uses GPU->Graphics interop, the performance in fps is extremely sensitive to anything that might slow the graphics down, like GPU->Host->GPU copies on the 2-GPU system.

I would use the BlackScholes SDK program as an independent test of the speed of CUDA calculations.

Oh, and one more idea: check that the GPU fans are clear and spinning and that the GPU temperatures are normal (50-60 C). An overheating GPU will cutperformance to save itself.

3

Please note that the CUDA release notes explain why Vista is slower than other operating systems.