Dual-GPU system is very slow: driver problem?
Dear comrades,

I have an i7 @ 3,7GHz and dual gtx-580, but only one is used to the computation. The OS is archlinux, CUDA is the newest, 4.2, gcc is 4.6. 294.40 is the driver.

My second computer is a E6600 @2,4Ghz, GTS 450. Cuda 4.1, gcc 4.4. 295.20.

The issue is: the same code is running faster in my humble computer. The time is 1 second. In my good computer, the time is 1,434 seconds. Even if a compile the code on the humble one and send to the good one, the time is the same.


The second issue: the kernel is four times faster in the good one, but the CPU time is 7x longer in the good one. In march this code used to return me 0,6 seconds. What should it be? A driver problem? Dual-gpu issue?

Thanks!

Tiago Carneiro
Dear comrades,



I have an i7 @ 3,7GHz and dual gtx-580, but only one is used to the computation. The OS is archlinux, CUDA is the newest, 4.2, gcc is 4.6. 294.40 is the driver.



My second computer is a E6600 @2,4Ghz, GTS 450. Cuda 4.1, gcc 4.4. 295.20.



The issue is: the same code is running faster in my humble computer. The time is 1 second. In my good computer, the time is 1,434 seconds. Even if a compile the code on the humble one and send to the good one, the time is the same.





The second issue: the kernel is four times faster in the good one, but the CPU time is 7x longer in the good one. In march this code used to return me 0,6 seconds. What should it be? A driver problem? Dual-gpu issue?



Thanks!



Tiago Carneiro

#1
Posted 04/23/2012 08:44 AM   
[quote name='CarcaraH' date='23 April 2012 - 10:44 AM' timestamp='1335170686' post='1399738']
What should it be? A driver problem? Dual-gpu issue?
[/quote]

At least update drivers to the latest version 295.41, and both to GCC4.6, then benchmark again. Anything else would be subject to speculation.
[quote name='CarcaraH' date='23 April 2012 - 10:44 AM' timestamp='1335170686' post='1399738']

What should it be? A driver problem? Dual-gpu issue?





At least update drivers to the latest version 295.41, and both to GCC4.6, then benchmark again. Anything else would be subject to speculation.

#2
Posted 04/23/2012 02:09 PM   
Or downgrade the newer gcc to 4.4, so you use a version supported by CUDA.
Or downgrade the newer gcc to 4.4, so you use a version supported by CUDA.

Always check return codes of CUDA calls for errors. Do not use __syncthreads() in conditional code unless the condition is guaranteed to evaluate identically for all threads of each block. Run your program under cuda-memcheck to detect stray memory accesses. If your kernel dies for larger problem sizes, it might exceed the runtime limit and trigger the watchdog timer.

#3
Posted 04/23/2012 03:56 PM   
[quote name='TZaman' date='23 April 2012 - 11:09 AM' timestamp='1335190157' post='1399853']
At least update drivers to the latest version 295.41, and both to GCC4.6, then benchmark again. Anything else would be subject to speculation.
[/quote]

unfortunately there is no 295.41 in archlinux repository.

And I tried to install gcc 4.4 using yourt, but with no success.
[quote name='TZaman' date='23 April 2012 - 11:09 AM' timestamp='1335190157' post='1399853']

At least update drivers to the latest version 295.41, and both to GCC4.6, then benchmark again. Anything else would be subject to speculation.





unfortunately there is no 295.41 in archlinux repository.



And I tried to install gcc 4.4 using yourt, but with no success.

#4
Posted 04/24/2012 03:50 PM   
Scroll To Top