On GLM-5.2/5.3, MI355X lands between Hopper and Blackwell. At 100 tok/s per user it serves ~26M tokens/s per MW, ahead of H200 (~16M) but behind B200 (~42M) and B300 (~44M), and its single-agent speed matches H200 at 228 tok/s, against 312 on B200 and 333 on B300.
