NVFP4 and MXFP4 multiply the same sixteen 4-bit values; the difference is one byte of scale per block.
We measured what that byte costs (41.5% of MXFP4 blocks clip their own maximum; about 9% of perplexity under plain rounding) and asked ptxas what eight NVIDIA targets do with each format.
No GPU needed, every number tiered by confidence.