Shuffling 100 million points per second (probably much more if it wasn't both, disc IO and CPU->GPU transfer limited), and shuffling in general if you want to maintain high throughput. e.g.: https://bit.ly/2zcUzJq

Monte Carlo simulations can have a significant speed dependency on the PRNG used.

That said, I'm not up to date on the speed of CSPRNGs so maybe they're already there.