Ancestors

Toot

Written by Robin Green on 2024-08-23 at 19:26

A new number representation enters the ring - Takums! All the increased precision of small values, less of the plummeting precision outside of the "normal" range.

https://arxiv.org/pdf/2408.10594

=> View attached media

=> More informations about this toot | More toots from fatlimey@mastodon.gamedev.place

Descendants

Written by aeva on 2024-08-23 at 19:37

@fatlimey what are the downsides

=> More informations about this toot | More toots from aeva@mastodon.gamedev.place

Written by Robin Green on 2024-08-23 at 20:36

@aeva This is literally the first paper I have heard of it, and not an endorsement. Just a note that it exists.

Analysis takes time, hardware takes longer. It's easy to invent number systems but a lot of work to prove them useful/fast/implementable, so I remain skeptical.

=> More informations about this toot | More toots from fatlimey@mastodon.gamedev.place

Written by Marc B. Reynolds on 2024-08-24 at 08:28

@fatlimey @aeva I like working with denormals but not as much as posits?

=> More informations about this toot | More toots from mbr@mastodon.gamedev.place

Written by Marc B. Reynolds on 2024-08-24 at 09:15

@fatlimey @aeva 4 drive-by readers: Tapered precision breaks the notion of scale invariance. So like posits you don't remove any issues related to finite representation and you add breaking scale invariance. So my rant (1) against posits applies (IMHO) just the log2 window sizes change. (also IMHO give more bits to log2 neighborhood of 1 isn't generally interesting). These are all trying to solve the wrong prob. I think the right problem is domain specific 16-bit.

1. https://marc-b-reynolds.github.io/math/2019/02/06/Posit1.html

=> More informations about this toot | More toots from mbr@mastodon.gamedev.place

Written by aeva on 2024-08-24 at 18:22

@mbr @fatlimey ty :D

=> More informations about this toot | More toots from aeva@mastodon.gamedev.place