AMD Instinct MI300X / MI325X / MI350X / MI350P / MI355X product specification tables (dense-vs-sparse convention)

Convention confirmed across five AMD pages on 2026-10-03 (mi300x, mi300a, mi325x, mi350x, mi350p, mi355x). AMD publishes TWO separate labelled rows per precision: 'Peak Half Precision (FP16) Performance 1.3 PFLOPs' = DENSE, and 'Peak Half Precision (FP16) Performance with Structured Sparsity 2.61 PFLOPs' = SPARSE, exactly 2x. Same pattern for TF32, FP8/OCP-FP8, INT8, INT4, MXFP4, MXFP6, MXFP8. Vector rows ('Peak Half Precision (FP16) Performance', 'Peak Single Precision (FP32) Performance') carry NO sparsity variant - scalar/vector paths on these parts have no 2:4 path, so those rows are dense by construction. Critically: the 'Specs Comparisons' chart on the SAME page prints e.g. 'FP16/BF16 Tensor / FP16/BF16 Matrix (Sparsity) 4.5 4.6' comparing MI350X to B200 SXM5 - i.e. AMD's own cross-vendor chart quotes B200's SPARSE figure against its own sparse figure, which is why an unaudited reader of that chart lands on the wrong dense/sparse reading for the NVIDIA part.