The question is, how much does your algorithm get from the 19 TFlops for a GPU, and how much from the 8 from the Versal. I'm sure many algos fit GPUs fine, but some don't, and might get more out of an FPGA.