NVIDIA Blackwell Dominates Agentic AI Benchmark

NVIDIA's Blackwell Ultra NVL72 platform shows leading performance on the new AgentPerf benchmark for agentic AI. It runs 20 times more agents per megawatt than NVIDIA Hopper systems. Here's the thing: Agentic AI works differently than conversational AI. Conversational AI is like a quick sprint, with one call and one response. But agentic AI is more like a relay race. It breaks a goal into many steps, chaining together dozens to hundreds of language model calls, along with tool calls like code execution and web browsing. This makes the complexity multiplicative. Existing AI benchmarks measure only a single language model call. They don't account for the chained calls, tool delays, and growing context that stress systems in agentic workloads. What's interesting is that the NVIDIA GB300 NVL72, using the DeepSeek V4 Pro model, delivers the highest performance in this new benchmark. It runs up to 20 times more agents per megawatt than the NVIDIA HGX H200 system. This advantage comes from extreme co-design across the entire system, connecting 72 GPUs into a single rack. The bottom line: Understanding these performance differences is crucial for companies building and deploying AI agents at scale.