tech ๐ค Nvidia crushes AMD in new AI agent benchmark ๐ค
SemiAnalysis released AgentX, an open-source benchmark testing real coding-agent sessions against production inference stacks. On GLM 5.3 via SGLang, Nvidia hardware showed up to five times better cost efficiency than AMD at 150 tokens per second per user. The B200 comparison against MI355X showed advantages nearing 57% and 247% at different speeds. SemiAnalysis argues that long-context multi-turn agentic sessions now dominate production inference traffic in the AI space. Buyers waiting for price drops from Nvidia are seeing the performance gap widen in this new testing framework.