
Nvidia Outpaces AMD in Agentic Workloads with Cost Efficiency Gains
The story in brief
On August 21, optimizations from Inferact and Nvidia merged upstream, enabling MI355X running SGLang to match B200 running vLLM. Nvidia's hardware showed up to five times better cost efficiency than AMD at 150 tokens per second per user. SemiAnalysis published AgentX on August 24, an open-source benchmark testing real coding-agent sessions. Nvidia's Rubin generation is set for later in August, with AMD's MI455X expected in 2026.