MXNorm: Reusing MXFP Block Scales for Efficient Tensor Normalisation

arXiv 2603.13180·medium signal

MXFP (microscaling floating point) quantization already computes block-level statistical summaries as a byproduct of quantization — MXNorm reuses these statistics for layer normalization instead of running a separate normalization pass. Eliminates redundant memory bandwidth consumption for normalization ops in quantized transformer inference without accuracy loss. Composable with existing MXFP inference stacks; no architectural changes required.