NVIDIA's GB300 NVL72 rack-scale Blackwell Ultra platform delivered the highest performance on the AgentPerf benchmark, showing up to 20x more agents per megawatt than the HGX H200 system.
NVIDIA's GB300 NVL72 rack-scale Blackwell Ultra platform delivered the highest performance in the first AgentPerf agentic AI benchmark, running up to 20x more agents per megawatt than the NVIDIA HGX H200 system.
AgentPerf, published by Artificial Analysis as the industry's first benchmark for agentic AI, measures chained model calls and tool-invocation patterns rather than single-request inference, reflecting how agents break goals into many steps and sustain long, context-rich interactions.
The benchmark runs DeepSeek V4 Pro, a large mixture-of-experts model, and reports how many concurrent agentic coding trajectories a platform can support while meeting defined thresholds for responsiveness and output token rate, with tool calls simulated as representative CPU processing time so results isolate accelerated compute performance.
NVIDIA credits the GB300 NVL72 advantage to full-stack codesign: 72 GPUs in a single rack-scale system, CUDA kernels that overlap communication and compute so coordination cost is absorbed, and TensorRT LLM optimizations that separate input processing from output generation to sustain efficiency as concurrent sessions scale.
Inference providers Baseten, DeepInfra and Together AI are already serving agentic workloads on Blackwell, including Together's real-time Cursor coding agents and DeepInfra's Pam.ai dealership agents running on the platform.
The company also notes its Vera Rubin architecture is now in full production, bringing additional infrastructure capacity aimed at growing agentic AI demand.
