Speaker
Dr
John Romein
(Stichting ASTRON)
Description
In this talk, I will report on a novel way to implement W-projection gridding on GPUs.
This method is several times faster than existing GPU gridders, because it needs significantly fewer accesses to the relatively slow GPU memory.
I will show how gridding on an eight-GPU system is typically 250 times faster than on a regular dual-CPU server, and over 20 times more energy-efficient.
The algorithm is programmed in both CUDA and OpenCL, which I will briefly compare.
I will also compare and explain the enormous performance differences between GPUs from Nvidia and AMD.
Primary author
Dr
John Romein
(Stichting ASTRON)
