cuFFT and Xavier (Solved)

This is probably a silly question but will there be an accelerated version of the cuFFT libraries for the Xavier that uses the tensor cores? From my little understanding the tensor cores seem to be a glorified quad MAC engine so could be used for that. My prime interest is in Software Defined Radio rather than AI although I have heard of AI being used in cognitive radio systems.

I know that Fairwaves have managed to get their XTRX product to work with a Jetson TX1 so I am hoping it won’t be too difficult to get it to work with the Xavier.

  • Charles

Hi,

Thanks for your question.

Tensor core use INT8 data format.
Currently, cuFFT can process half-precision data input but not for INT8 yet.

You can check if your software can benefit from fp16 acceleration first.
[url]https://docs.nvidia.com/cuda/cufft/index.html#half-precision-transforms[/url]

Thanks.

Thanks obviously I had misunderstood this link
https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/

My main question was does the Xavier support cuFFT as I see no mention of it in the Jetpack documentation I have read so far.

  • Charles

Hi Charles, yes, cuFFT library is included with the CUDA toolkit and JetPack for Jetson AGX Xavier.