IEEE P3109 Standard defines efficient ML arithmetic with stochastic rounding

The IEEE P3109 draft standard, introduced in an arXiv paper by Andrew Fitzgibbon, Christoph M. Wintersteiger (both Microsoft Research), and Jeffrey Sarnoff, defines a flexible family of binary floating-point formats purpose-built for machine learning. Unlike general-purpose IEEE 754, P3109 formats are parameterized over bit width, precision, signedness, and the presence of infinities. All operations decode values into the closed extended reals (reals, ±infinity, NaN) to ensure only real arithmetic is invoked—eliminating hardware exceptions and accelerating throughput. The standard introduces stochastic rounding as a core mode, critical for low-precision training. A novel scale-invariant error metric, 'kappa-approximation,' allows vendors to describe approximate implementations with bounding analogous to units in the last place.

The practical impact of P3109 is significant for ML hardware and software designers. By offering exception-free operations and explicit handling of NaN/infinite operands through return values, the standard streamlines accelerator design and reduces latency. The parameterized formats enable single-chip support for varying precision needs (e.g., 4-bit inference vs. 8-bit training) without redundant logic. Block operations with shared scale factors are uniformly defined, simplifying matrix math. The standard also includes mechanical verification via formal specifications, ensuring correctness across implementations. This work positions P3109 as a foundational arithmetic layer for next-generation AI accelerators, balancing efficiency, consistency, and flexibility.