ailocalinference

← Glossary

NVFP4 (Blackwell FP4)

NVIDIA's hardware 4-bit floating point format with micro-block scaling

Formats

In plain English

A micro-engineered 4-bit number system that lets Blackwell GPUs pack 2x more AI power into the same silicone memory.

Technically

NVIDIA's 4-bit floating point format (E2M1) introduced on Blackwell Tensor Cores. Groups 16 elements into micro-blocks sharing an FP8 scale factor, delivering 2x matrix multiplication throughput over FP8 with near-zero perplexity degradation.

Related terms

QuantizationGB10 SuperchipDGX Spark

More in Formats

AWQBitNet b1.58 (1-bit LLMs)GGUFGPTQQuantization