Comment by magicalhippo

> Everything that can run in parallel, will run in parallel.

On the CPU, there's typically a threshold where dividing and coordinating the parallel work takes more time than simply doing the work on a single thread. Thus you can make the overall runtime much faster by not dividing the work all the way, but rather stop at that optimal threshold and then just loop over the remaining work in the worker threads.

How does this work on the GPU using Bend? Been too long since I did any GPU programming.