positing variable precision

rob mahurin — Sun, 07 Sep 2025 05:01:36 GMT

From this year's "summer of math exposition" collection , an explainer on the posit, a variable-precision representation of floating-point numbers.

The posit uses fewer bits to encode small exponents, so numbers whose magnitude is roughly unity can get a few extra bits of precision in the mantissa.

The intuition that most floating-point values have magnitude roughly unity is certainly consistent with my experience, and also consistent with guidance that I have received from various computing mentors. But it would be interesting to build a virtual CPU and monitor which values enter the FPU during normal operation, to check.

mask of the robtasm (Posts about ieee754)

positing variable precision