Comment by exikyut

Hmmmmm.

Reading the aside

In fact IEEE floating-point math is designed to, whenever practical, give the best possible answer (correctly rounded)

Made me realize - there are not "just" four billion floats, there are four billion floats multiplied by however many levels of rounding you want to test against.

The "ninety seconds" doesn't qualify if the test harness was multithreaded. If it wasn't, a modern >30 core box (maybe buried in CI somewhere) should be able to manage quite a few rounding levels without difficulty thankfully.