Trouble is, you tend to have to use realtime friendly CPU's when you do this stuff. So things like pipeline optimization, predictive branching, cache management, etc, tend to not always be present on CPUs that are designed for determinism. So realistically, I would not be surprised if the heaviest computing in that thing is done on a single core 500MHz-1GHz CPU, and the rockets aren't probably running anything fancier than a 100 MHz single core machine either. They may or may not have memory management in hardware (I know some higher end 1GHz SBCs I've worked with don't, they just have a single flat address space).
It's still way overkill for what it needs to do, but in order to get the cycle times down under the 10 msec mark (order of magnitude), you have to pick and chose what language you work in (anything with non determinism like Java or something with a heap like the C++ run time may well be a no-no), and you have to pick and choose which algorithms you implement.
The latest-and-greatest randomized algorithm you see in IEEE may not be suitable for realtime implementation because while the average time is 1msec, there exist execution paths which take 15 msec to run. So you're safer going with the sub-optimal thing that executes in 5 msec every time.