MXFP4
The OCP microscaling 4-bit format: 32-value blocks sharing a power-of-two scale. NVIDIA's NVFP4 uses 16-value blocks with an FP8 scale, and the choice is baked into silicon and checkpoints.
Current numbers
288 GBMI355X HBM3E capacity at 8 TB/s, ~10.1 PFLOPS dense MXFP4, 1,400 W DLC (CDNA 4, N3P)
~10.1 PFLOPSAMD MI355X MXFP4 peak; AMD lists no sparse figure, and MXFP6 runs at the same rate on CDNA4
16 vs 32NVFP4 block size (16 values, FP8 scale) vs MXFP4 (32 values, power-of-two scale)