AI artificial intelligence is a hot spot in recent years，Both CPU and GPU are improving AI performance，Intel, which was originally a competitor 、 NVIDIA and ARM also team up for the first time，FP8 format launched ，This new AI format can improve computing performance，The performance of the current strongest accelerator card H100 has been improved by 4.5 times。
Three companies join forces to launch common AI format related to current chaos，Many people know that FP32 single precision is often used to measure GPU performance before.、FP64 double precision format，However, in the age of AI, such high precision is not required，In addition, manufacturers often customize the calculation format，So INT8 has also appeared in recent years、BF16、AI standards like TF32。
Manufacturers fighting on their own is obviously not conducive to the development of AI technology，So Intel、The three major CPU/GPU giants of NVIDIA and ARM have jointly launched the AI standard format FP8，And released the FP8 Formats for Deep Learning white paper，Also submitted the standard to the IEEE Society，If nothing else, it will become a general standard in the field of AI.。
According to them，The FP8 standard will have two encoding variants，E5M2 respectively、E4M3，Deviations from the current IEEE754 standard can be minimized，Not only can it bring higher flexibility to AI computing，Balance between hardware and software，Improve developer productivity。
According to the results published by NVIDIA，FP8 has 16-bit floating point-like effects in widely used training models，and better performance，NVIDIA's H100 accelerator card uses the new FP8 format，Achieved 4.5x performance improvement in BERT model。
It is worth noting that，Three major manufacturers jointly define the FP8 standard，But did not see the figure of AMD，I don't know what AMD's attitude is，However, AMD should eventually adopt the FP8 standard，unify with the three。
