Technical details on what I managed to accomplish:
While poking around in the USML codebase I figured out there are faster but more approximate ways to do some of the things that I was doing, so using those I managed to get the average time between updates down to 0.7 seconds per contact at 20nm spread across all cores. Instead of replacing the linear algebra library, I decided to eliminate one very expensive matrix operation and replace it with a bunch of individual scalar operations. That brought computation time per contact down to 0.5 seconds. I'm thinking that's adequate.
These computation times are calculated based on the average time for 4 separate benchmarks to run to completion divided by 4. There's a little more to it than that, but it's important to note a single computation actually takes around 2 seconds, but there are 4 happening at once so you'll get one update every 0.5 seconds on average.
So now all that remains is creating the actual game.
