SemiAnalysis reports AMD MI355 runs 1.42x faster than GB300 NVL72 on TileRT

SemiAnalysis says AMD's MI355 is 1.42x faster than Nvidia's GB300 NVL72 on long-context inference when running the TileRT software, which it says reaches 1330 p90 tokens per second per user.

The claim is posted on X by SemiAnalysis, and the post credits AMD leadership for providing an MI350 cluster to TileRT maintainers for the test.

The post does not include the full benchmark setup or the workload details beyond long context, so the comparison cannot be checked from the report alone.