The MI300X, MI325X and MI355X support FP8. The MI250X and MI100 do not. The two formats differ by generation, which matters when you move FP8 models between chips.
The support table
From AMD's ROCm precision support page, for the HIP C++ FP8 types:
| GPU | Architecture | FP8 format supported |
|---|---|---|
| MI355X | CDNA4 | OCP (e4m3, e5m2) |
| MI325X | CDNA3 | FNUZ (e4m3, e5m2) |
| MI300X | CDNA3 | FNUZ (e4m3, e5m2) |
| MI250X | CDNA2 | None |
| MI100 | CDNA1 | None |
The matrix-core tables on the same page show float8 (E4M3 and E5M2) supported on the CDNA3 and CDNA4 parts and not on CDNA2 or CDNA1. AMD also notes that library support depends on hardware support, so an FP8-capable library only works fully on the architectures marked above.
Why it matters
FP8 checkpoints built for one format may need conversion for the other. If you are weighing these chips against NVIDIA's, start from the specs in the MI325X guide and the MI355X guide, and see NVIDIA's main competitors in AI for the wider field.
