slow FP16 cuFFT

srawrats October 8, 2018, 4:49pm 1

I’m trying to do an FFT on a device with compute capability 7.5 using FP16 inputs and outputs. I’ve noticed that, using the profiler:

1k FFT (complex float): 2.58 us
1k FFT (complex half-float): 2.68 us

So I’m curious if cuFFT has been optimized for half-floats.

Topic		Replies	Views
FFT performance loss? FFT-Size 16k and above CUDA Programming and Performance	0	2145	May 9, 2008
Cufft2d FP16 and BF16 is slower than FP32 GPU-Accelerated Libraries cufft	1	682	June 9, 2023
FP16 and CUFFT GPU-Accelerated Libraries	0	1176	July 27, 2015
Unexpectedly low performance of cuFFT with half floating point (FP16) GPU-Accelerated Libraries	1	1661	June 16, 2017
cuFFT doubt. GPU-Accelerated Libraries	1	808	January 18, 2015
CuFFT FP16 is slower that FP32 Jetson Xavier NX cuda	5	1199	April 5, 2023
Half precision cuFFT Transforms GPU-Accelerated Libraries	12	6073	March 29, 2021
CUDA 9 cufft slower for some sizes GPU-Accelerated Libraries	3	873	October 1, 2017
double precision CUFFT 2.3 is to slow CUDA Programming and Performance	2	5066	August 26, 2009
cuFFT DFT Performance question GPU-Accelerated Libraries	1	517	November 20, 2019