AIFloats

The IEEE SA P3109 Working Group is drafting a standard for floating-point formats used in machine learning.

We are defining a family of data formats and providing binary arithmetic optimized for machine learning and related domains. The intent is to provide a consistent and flexible arithmetic framework optimized for Machine Learning Systems that may be implemented in hardware or in software or both combined. Having a shared standard improves interoperability. The formats are designed to be efficient, consistent and highly performant.

Here is the source code for this package.