MI355X has its strongest showing on DeepSeek V4 Pro 0813. It serves ~15% to 40% more throughput per MW than B200 between 35 and 75 tok/s per user, peaking at about 94M tokens/s per MW at 33 tok/s per user, a level B200 only reaches at 13 tok/s. Above about 80 tok/s per user B200 pulls ahead, and for a single agent the gap becomes MI355X at 240 tok/s vs. B200 at 367 tok/s. Disaggregated multi-node configs for other NVIDIA hardware largely sit above both, e.g. with B300 at 125M tokens/s per MW at 116 tok/s per user using two 8-GPU nodes.
