Computing requirements for artificial intelligence are increasing exponentially. New innovations across hardware and software are needed to achieve the computing throughput required to advance AI.

Intel announced on the 16th that it has co-authored a paper with ARM and Nvidia describing the 8-bit floating-point (FP8) specification and the E5M2 and E4M3 to provide an interchangeable format for artificial intelligence (AI) training and inference. Through this inter-company agreement on specifications, it is expected that various AI models will be able to operate and perform consistently across hardware platforms, thereby accelerating AI software development.

One of the research areas gaining prominence to address the growing computing gap is reducing numerical precision requirements in the field of deep learning to improve memory and computational efficiency. Precision reduction methods enhance computing efficiency by leveraging the inherent noise recovery properties of deep neural networks.

Intel announced that it plans to support the specifications in its AI product roadmap for CPUs, GPUs, and other AI accelerators, including the Havana Gaudi deep learning accelerator.

FP8 minimizes deviations from the existing IEEE 754 floating-point format based on a balance between hardware and software, utilizing existing applications, accelerating new adoption, and improving developer productivity.

The principle of the format proposals by ARM, Nvidia, and Intel is to utilize protocols, concepts, and algorithms based on IEEE standardization. This format proposal is expected to grant the greatest degree of freedom for future AI innovation while adhering to current industry protocols.