Futhark 0.6.1 released – High-performance functional programming on the GPU

It depends on the data and the workload. I did some work on GPU accelerating viewshed calculations - basically, line of sight from a point where cells are marked visible or not visible. It's really useful in radio mast planning.

In this case, the result of any given cell doesn't rely on the result of any other cell. This leads to the neat case where every single cell can be, in theory, calculated at once. In this case, you essentially get n-times speedup for n-time increase in processing power.

The other, often overlooked, bonus is that the GPU simply isn't the CPU. As long as you're waiting on stuff to finish there, your CPU is free to do what it pleases. In my case, whenever I started, I immediately triggered the intialisation of the GPU, and then started reading command line switches, reading in the data from disk, etc. Likewise, when the GPU was doing the heavy lifting, I could start on the slow IO involved in preparing the output, creating files, writing out metadata about the shape and limits of the output that's trivially calculated.

In the end, the speedup varied, but at peak it was turning an 8-hour CPU bound workload into one that taken 2 minutes. What was once a full days cycle where you set up your simulation, did busywork for the day, and then collected results and thought about the result overnight at home can now be done while getting a coffee, and allowed people to make mistakes and play around.