Tesla C1060 and GTX 275 running same code on two platforms

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, all,

In this summer, I wrote a CUDA GPU linear solver and tested it on Tesla C1060. It performed well on it.

Now I copy the same code and recompile and run it on GTX 275.

The result of it is very strange for me

the output is not consistent for each run. It means if I run the code several times. The solution may vary. Usually it varies a little. BUT sometimes, it gives me a NAN value.

I’m sure the code is good on Tesla C1060 since we test it for a whole summer. I use Cublas in the code.

Can anybody give me a guess? Where may the problem be? The code itself? The compilation? The platform?

Thanks!

2

Race condition due to different clocks?

3

I would guess it is a watchdog timer issue. Are you running a display on your GTX 275?

4

Problem is in the “code”. :-)

As Tim suggested, the different clock rate has exposed the bug… Thats all.

One can hope it is not from cuBLAS as it is tested in lot of different configurations…