4-bit floating point FP4

Article

• Explains FP4 floating point format: 1 sign, 2 exponent, 1 mantissa bit
• Covers trade-offs: no NaN/Inf support in basic variants; very limited precision
• Relevant to LLM quantization (NVFP4, MXFP4 formats)

Discussion

• Commenters note FP4 lacks NaN/Inf; suggest palette/lookup-table approach instead
• Discussion of minifloats history and x87 80-bit legacy
• One commenter notes this was an April Fools joke format nine years ago

Discuss on HN