NVIDIA Blackwell Takes First Place in the First Agentic AI Infrastructure Benchmark

According to the results of AgentPerf, the first agentic AI benchmark released by Artificial Analysis, the NVIDIA Blackwell Ultra NVL72 platform achieved leading performance across all tested agentic AI workloads. In particular, the NVIDIA GB300 NVL72 demonstrated efficiency that can drive up to 20 times more agents per watt than NVIDIA HGX H200 systems.

Unlike traditional conversational AI, which responds with a single LLM call, agentic AI is a relay-style workload that repeats multiple stages of LLM calls and tool calls to achieve a goal. In this process, context grows exponentially and computational complexity multiplies, making it difficult to measure real-world performance with existing single-call-based benchmarks.

NVIDIA's performance advantage stems from full-stack co-design:

• GB300 NVL72: Connects 72 GPUs into a single rack-scale system, efficiently distributing the execution of large-scale MoE models such as DeepSeek V4 Pro.
• CUDA kernels: Overlap communication and computation so that coordination costs between experts are absorbed without adding to latency.
• NVIDIA TensorRT LLM: Separates input processing from output generation, maintaining efficiency even as concurrent agent sessions scale.

AgentPerf is designed based on real coding agent trajectories. It uses data from real public code repositories extracted across 12 or more programming languages, reflecting actual workflows in which agents read files, write and modify code, execute commands, and iterate.