Cross-industry hardware specification to accelerate AI software development

Cross-industry hardware specification to accelerate AI software development

Arm, Intel, and Nvidia jointly wrote a paper describing an 8-bit floating point (FP8) specification and its two variants E5M2 and E4M3 to provide a common interchangeable format for artificial intelligence (AI) and distraction work. This cross-industry specification alignment allows AI models to operate and function consistently across all hardware platforms, accelerating AI software development.

Computing requirements for AI have grown exponentially. Innovations in hardware and software are needed to provide the processing power needed to advance AI.

One of the promising areas of research to address this growing computational gap is to reduce the numerical accuracy requirements for deep learning to improve memory and computational efficiency. Low-precision methods take advantage of the inherent noise resilience properties of deep neural networks to improve computational efficiency.

Intel plans to support this format specification in the AI ​​product roadmap for CPU, GPU, and other AI accelerators, including Habana Gaudi deep learning accelerators.

FP8 minimizes deviations from existing IEEE 754 floating point formats with a good balance of hardware and software to leverage existing implementations, accelerate adoption and improve developer productivity.

The guiding principle of this format proposal by Arm, Intel, and Nvidia is to use conventions, concepts, and algorithms built into the IEEE standardization. This provides the broadest scope for future AI innovations in line with current industry conventions.

Translate »