Nelder-Mead is used in GPU

I was wondering how fast do you think this algorithm would run on GPU under the following condition:

  • Trying to optimize a 3d.
  • The function we try to optimize takes significantly more time to run than the Nedler-Mead.

From what I have seen, Nelder-Mead implementations have lots of conditional branches. Will thus hurt my performance if i run a nelder-mead kernel for many points? (i.e. run the optimization function when it is not needed?) Is there way I can do this quickly? Would it be wise to do the nelder-mead iterations and comparisons on CPU and do only the calculations of the optimization function in GPU.

Rob Farber’s new book may help: Farber 2011. Pages 36-61 explain and implement Nelder-Mead.