AMD MI355X reportedly narrows performance-to-cost gap with GB300 in AI-agent inference

SemiAnalysis reports that AMD’s MI355X is quickly closing the gap with GB300 in performance relative to total cost of ownership for agentic inference. It attributes the improvements to AMD’s SGLang team and MoRI (Modular RDMA Interface) library, and says they also reflect UMBP’s integration into SGLang as a KV offloading backend.