The first infrastructure benchmark for agentic AI, called AgentPerf, has been introduced by @ArtificialAnlys, providing a means for developers and enterprises to evaluate accelerated computing systems designed for these advanced applications. This new benchmark addresses the limitations of existing metrics by focusing on the chained model calls, context gathering, and iterative processes that characterize agentic workloads. Initial results indicate that NVIDIA Blackwell significantly outperforms NVIDIA Hopper, delivering 20 times more agents per megawatt, underscoring NVIDIA’s commitment to optimizing hardware and software performance for agentic AI reasoning tasks.

NVIDIA: NVIDIA develops GPUs and full-stack AI computing platforms used for training and inference. Its Blackwell architecture is highlighted in the news for delivering superior efficiency in agentic AI tasks compared to prior generations like Hopper. The company actively collaborates on real-world benchmarks for emerging AI workloads.
AgentPerf: AgentPerf, also known as AA-AgentPerf, is a hardware benchmark introduced by Artificial Analysis to measure inference system capacity under realistic agentic AI conditions. It focuses on supporting multiple concurrent agents through chained model interactions rather than isolated queries. The benchmark enables direct comparisons of accelerated computing platforms for this new class of workloads.
Artificial Analysis: Artificial Analysis provides independent benchmarks and analysis for AI models, APIs, and infrastructure providers. It recently launched AgentPerf as the first benchmark specifically designed to evaluate hardware performance on agentic AI workloads involving chained model calls and tool use. The organization supports developers, enterprises, and infrastructure providers in comparing accelerated computing systems.

Agentic Workloads: Agentic AI applications increasingly rely on chained model calls, context gathering, and iterative tool use, creating demand for specialized infrastructure benchmarks.
Infrastructure Focus: NVIDIA is emphasizing full-stack co-design across hardware, interconnects, and software to optimize performance for agentic AI reasoning tasks.