FP8
An 8-bit number format, supported since Nvidia's Hopper GPUs in 2022, that doubles AI speed over 16-bit formats.
Also: 8-bit floating point
FP8 stores each number in 8 bits, in two common variants that trade range against detail. Nvidia's Hopper generation added hardware support and a "Transformer Engine" that chooses between precisions layer by layer, and AMD's MI300 series also supports it. FP8 is widely used for inference and, increasingly, for training: DeepSeek reported training its V3 model largely in FP8 in late 2024. Using it well needs scaling factors so that values do not overflow the small range.