Running executables on different machines Quadro FX570 vs Tesla

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all,

If I compile my code with nvcc on a fedora 9 box that has a quadro fx570 and then transfer and run the executable on a fedora 8 machine that has a Tesla server attached can I expect the executable to run at full speed???

I guess I’m asking does nvcc compile for my quadro FX570 that has 2 MPs adn so will only use 2 MPs on the tesla card???

I see about 3 times increase in peformance when running on the tesla compaired to my quadro. I was expecting a little more. I haven’t multiGPU’d the code so I was expecting roughly a 8x (16MP/2MP) performance increase.

My kernel launches around 37K threads so I guess this would be enough to fill the tesla’s 128 scalar processors (on 1 GPU)???

Thanks,

Wes.

2

You have to launch many blocks. Blocks are the ones that get scheduled across MPs and will let your application scale.

3

Yes, the executable doesn’t need to be recompiled to adapt to new hardware. Your program will use as many MPs as the card has, provided you have enough blocks to saturate them.

Maybe try running SDK examples on the two machines and see how they sped up?