NVIDIA Blackwell Leads First Agentic AI Infrastructure Benchmark
Original: NVIDIA Blackwell Leads on First Agentic AI Infrastructure Benchmark
NVIDIA says GB300 NVL72 tops AgentPerf, delivering up to 20x more agents per megawatt than HGX H200.
NVIDIA reports that its GB300 NVL72 platform leads the first published AgentPerf results from Artificial Analysis, a benchmark designed for agentic AI infrastructure. The benchmark uses DeepSeek V4 Pro and coding-agent-style workloads with long sequences, simulated tool delays, and concurrency targets. NVIDIA attributes the gains to rack-scale Blackwell design, CUDA optimizations, and TensorRT LLM, claiming up to 20x more agents per megawatt than HGX H200.
NVIDIA’s blog post presents the first published results from AgentPerf, an agentic AI infrastructure benchmark created by Artificial Analysis, and says the NVIDIA GB300 NVL72 platform delivered leading performance across the tested workloads. The central claim is that GB300 NVL72 can run up to 20x more agents per megawatt than an NVIDIA HGX H200 system on the benchmark’s DeepSeek V4 Pro workload. NVIDIA frames this as a sign that agentic AI needs a different kind of performance measurement than traditional single-request LLM inference benchmarks.
Free shows the 3-line summary; Pro unlocks the full deep summary (~300 words) so you never have to click through.
See Pro plans →Want the original English / full article?
Read on NVIDIA Blog →Related
Summaries are AI-generated; the original article is authoritative.