LeetArxiv

LeetArxiv

Share this post

LeetArxiv
LeetArxiv
Mediant32 : An Alternative to FP32 and BF16 for Error-Aware Compute

Mediant32 : An Alternative to FP32 and BF16 for Error-Aware Compute

Computer Architecture : Replacing Floating Point Arithmetic with Integer Fractions on the Stern-Brocot tree

Murage Kibicho's avatar
Murage Kibicho
Apr 09, 2025
∙ Paid
5

Share this post

LeetArxiv
LeetArxiv
Mediant32 : An Alternative to FP32 and BF16 for Error-Aware Compute
1
Share
This article is part of a series on experimental computer architectures.
Related Articles
  1. Stern Brocot Fractions as a floating point alternative

  2. Emulating a GPU on a CPU using Finite Field Theory

Subscribe for updates

Features of the Mediant32 Number System

In our previous article, What every programmer should know about Stern Brocot Fractions, we introduced the Stern-Brocot tree as an alternative to fixed-point and floating-point arithmetic for representing fractions on parallel computers for AI.

As promised, this article explores addition, subtraction, multiplication and division on the Stern-Brocot tree.

We call this the Mediant32 number system. Further on, we proceed to implement fundamental arithmetic operations in this number system.

Mediant32 is an experimental number system for efficient AI compute. Sign up for updates below.

Keep reading with a 7-day free trial

Subscribe to LeetArxiv to keep reading this post and get 7 days of free access to the full post archives.

Already a paid subscriber? Sign in
© 2025 Murage Kibicho
Privacy ∙ Terms ∙ Collection notice
Start writingGet the app
Substack is the home for great culture

Share