Arm Intel NVIDIA publish OCP OFP8 FP8 formats

Arm, Intel, and NVIDIA have published an OCP 8-bit Floating Point Specification to establish standard interchange formats for machine learning inference and training. The FP8 family consists of the E4M3 and E5M2 encodings, balancing precision and dynamic range differently to suit various tensor roles. Additionally, Arm has introduced the Floating-Point Mode Register and new instruction sets across its architecture to make FP8 data types practical on CPUs. Open article →