P3109 - Arithmetic Formats for Machine Learning

IEEE P3109 – Arithmetic Formats for Machine Learning

Scope of proposed standard: This standard defines a binary arithmetic and data format for machine learning-optimized domains. It also specifies the default handling of exceptions occurring in this arithmetic. This standard provides a consistent and flexible arithmetic framework optimized for Machine Learning Systems (MLSs) in hardware and/or software implementations to minimize the work required to make MLSs interoperable with each other as well as other dependent systems. This standard is aligned with the IEEE Std 754-2019 Standard for Floating-Point Arithmetic.

Need for the Project: Machine Learning Systems have different arithmetic requirements from most other domains. Precisions tend to be lower, and accuracy is measured in dimensions other than just numerical (e.g. inference accuracy). Furthermore, Machine Learning Systems often are integrated into mission-critical and safety-critical systems. With no standards specifically addressing these needs, Machine Learning Systems are built with inconsistent expectations and assumptions that hinder the compatibility and reuse of machine learning hardware, software, and training data.

Stakeholders for the Standard: System developers, vendors and users of machine learning applications across many industries and interests including but not limited to compute, storage, medical, telecommunications, e-commerce, fleet-management, automotive, robotics, and security.


To express interest in joining this working group, please submit your contact information and follow the instructions here.


Supporting documentation:

Project Authorization Request (PAR)