Arm, Intel, and Nvidia have jointly authored a paper describing an 8-bit floating point (FP8) specification and its two variants E5M2 and E4M3. This should provide a common interchangeable format that works for both AI training and inference.

This cross-industry specification alignment will allow AI models to operate and perform consistently across hardware platforms, accelerating AI software development. Computational requirements for AI have been growing at an exponential rate. Innovation is required across hardware and software to deliver the computational throughput needed to advance AI.

One of the promising areas of research to address this growing compute gap is to reduce the numeric precision requirements for deep learning to improve memory and computational efficiencies. Reduced-precision methods exploit the inherent noise-resilient properties of deep neural networks to improve compute efficiency.

FP8 minimizes deviations from existing IEEE 754 floating point formats with a good balance between hardware and software to leverage existing implementations, accelerate adoption and improve developer productivity. The guiding principle of this format proposal from Arm, Intel, and Nvidia is to leverage conventions, concepts, and algorithms built on IEEE standardization. This enables the greatest latitude for future AI innovation while still adhering to current industry conventions.