Jetson Nano cuFFT

Autonomous Machines Jetson & Embedded Systems Jetson Nano
#1

Hello,

I have a question regarding cuFFT computed on Jetson Nano. I need to compute 8192 point FFT 200000x per socond. Is there anybody who has experience with Jetson Nano and cuFFT? Does the Jetson Nano have enough power to compute it?

Thank you for your support.

Martin

#2

Hi,

You can find a sample in /usr/local/cuda-10.0/samples/7_CUDALibraries/simpleCUFFT.

We update the example to 8192 signal and kernel size = 11.
The whole application execution time is

real	0m0.653s
user	0m0.456s
sys	0m0.152s

If 200000 signal with kernel size=11, the execution time is

real	0m0.974s
user	0m0.712s
sys	0m0.212s

Thanks.

#3

Hello,

thank you for your support, but I do not know if I understand correctly. Does the first table mean 8192 points fft which is calculated 2000000times? Does it mean that calculation time is 456ms? It means that Jetson nano is able to calculate 8192 points fft more than 400000times per second. Does the second table mean 200000 points fft calculated 200000 times per second?

Thank you for your answer.

Martin

#4

Hi,

We don’t test it 200000 times. You can try it on your own.

The experiment above is just to give you some idea about the performance.
The real profiling data should also take the kernel size into account.

By the way, please noticed that the most time consuming stage is the memory allocation and memory copy.
So if you run the application N time, the expected performance should be much better than NxT.

Thanks.

#5