Actually, I had to use really huge floats in one of my projects once and I can imagine use for them other than counting things.
The task was using genetic algorithms in planning a schedule, with quite a few factors counted in. Whoever worked with genetic algorithms will know they work rather poorly on mostly flat quality functions but converge really nicely if the function is really very steep, so instead of sum of squares you can use product of squares. The solution can be found in 1/100th or less the time but the problem is that for quite a small set to be optimized you start at 1e-50 and end about 1e+100 "quality" values, with bigger sets the values grow rapidly - so you need to use bignum libraries, i.e. 256-byte float arithmetics. They create some serious overhead but profit of fast algorithm convergence beats the loss at software efficiency by far.

This is not a task for a '51 though.
Think DSP maybe?