Nice discussion on whether #halfprecision #floats are worth the effort or not, specifically using the #gpu https://futhark-lang.org/blog/2021-08-05.. Turns out the #speedup really isn't so much. I suspect it *may* be better when you have large vectors, you can literally keep twice as much in #cache and larger models in #ram ... But still, it's worth checking whether you actually get the expected speed-up or not.