Get the latest tech news
Posit floating point numbers: thin triangles and other tricks (2019)
First post in breaking down why the posits presentations are egregiously disingenuous.
In the IEEE case the relative error is flat and the pair of plots (Kahan’s method and exact correctly rounded) are virtually identical excluding the ends of the usable range. For posits the relative error (if you flip it upside down) looks like rolling a ball down a ziggurat and worse there’s generally a large gap (remember log scale) between the exact correctly rounded and Kahan’s method results. A bottom line here is if you want to improve performance and/or accuracy of some computations…well if someone hasn’t already done the work for you then some math (algebra or otherwise which you may or may not find tedious) is required and writing any code is an error prone process (but somebody’s gotta do it).
Or read this on Hacker News
