Using cuFFT with `legacy’ callbacks (the only kind of callback supported before 12.6), we observe a performance decrease when moving from CUDA 11.7 to CUDA 12.2 or 12.4 (12.6 is not available on the HPC platform in question). This performance decrease is approximately 20%, and the observed behaviou…

Thanks. crossposted to GPU Accelerated libraries for the moment. I’ll file a proper bug when I can reproduce this without our full HPC code.

Performance degradation from CUDA 11.7 to 12.2 and newer with cuFFT

rs277 November 12, 2024, 8:13pm 2

You may get assistence on the GPU Accelerated Libraries forum or failing that, filing a bug.

Topic		Replies	Views
Performance degradation from CUDA 11.7 to 12.2 and newer with cuFFT GPU-Accelerated Libraries cufft	0	94	November 12, 2024
What is recommended for using cufft callbacks? GPU-Accelerated Libraries cufft	0	489	December 18, 2023
Performance Degradation in cuFFT callback from CUDA 11.8 to 12.9+ GPU-Accelerated Libraries cufft	0	63	November 3, 2025
cuFFT allocating and deallocating memory at every step GPU-Accelerated Libraries cufft	2	1054	February 5, 2023
cufftPlan creation deadly slow on CUDA 11+ GPU-Accelerated Libraries cufft	11	3772	February 1, 2023
What specifically is deprecated about cuFFT callbacks in CUDA 11.4? GPU-Accelerated Libraries cufft	6	1401	October 16, 2024
cuFFT callbacks slow GPU-Accelerated Libraries	4	949	August 15, 2018
Cufft load callback causes cufftExecC2C to allocate memory? GPU-Accelerated Libraries cufft	3	754	September 16, 2023
cuFFT callbacks not working with large amount of host-pinned memory? GPU-Accelerated Libraries	5	1098	July 6, 2017
cuFFT performance decrease with CUDA 7.0 GPU-Accelerated Libraries	2	860	March 27, 2015