FP4
A 4-bit number format introduced for AI with Nvidia's Blackwell GPUs, mainly used to make inference faster and cheaper.
Also: 4-bit floating point, NVFP4, MXFP4
FP4 stores each number in just 4 bits, so only 16 distinct values are possible. To make that workable, numbers are grouped into small blocks that share a scaling factor, a method called microscaling, defined by an Open Compute Project specification in 2023 (MXFP4) and in Nvidia's own NVFP4 variant. Blackwell GPUs run FP4 at twice their FP8 speed, which is where the largest headline performance figures come from. FP4 is used mostly for inference, with research under way on training in it.