AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing?

· SemiAnalysis

SemiAnalysis has released AgentX 1.0, the world's first fully open-source, multi-turn agentic coding inference benchmark designed for 1 million context. This benchmark addresses the rapid growth of long-context, multi-turn agentic workloads, which now dominate production inferencing traffic. AgentX 1.0 improves upon previous benchmarks by using agentic coding traffic instead of single-turn traffic, providing a more accurate measure of AI hardware and software performance. The benchmark runs on significant compute across various SKUs, including MI355X, GB300, B200, and H200. It highlights performance themes across frontier models like DeepSeek V4 Pro 0813, considering metrics such as performance per dollar, time to first token (TTFT), and overall task completion. The report notes strong performance from both NVIDIA and AMD on agentic workloads, with AMD's ATOM showing competitive results against NVIDIA's offerings in some scenarios. The project has already driven over 70 upstream PRs for optimizing real-world production agentic workloads across various frameworks.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.