Report
MI355X reportedly delivers 15%–40% more DeepSeek V4 Pro 0813 throughput per MW than B200 at 35–75 tokens/s per user
ArtificialAnlys says MI355X delivers roughly one-third of B300's throughput per MW on Kimi K3 at 50 tokens/s per user.
TLDR
ArtificialAnlys reports MI355X delivers roughly 15%–40% more DeepSeek V4 Pro 0813 throughput per MW than B200 at 35–75 tokens/s per user. B200 pulls ahead above about 80 tokens/s per user and reaches 367 tokens/s for a single agent, versus MI355X's 240. On Kimi K3, ArtificialAnlys puts MI355X at about 8 million tokens/s per MW at 50 tokens/s per user, versus about 24 million for B300 using a multi-node disaggregated setup.
Combined views
—
5 Sources, first seen ago
