Hardware·
NVIDIA Vera Rubin Platform Sweeps MLPerf Training and Endpoints Benchmarks
NVIDIA's new Vera Rubin platform delivers record‑breaking results across every MLPerf Training v6.1 benchmark and achieves over 10× generational throughput‑per‑watt gains in the inaugural MLPerf Endpoints submission, signaling a major leap for agentic AI workloads.

NVIDIA's Vera Rubin platform made a clean sweep of the MLPerf Training v6.1 suite, posting the only complete set of results across large language models, image generation, recommender systems and reinforcement‑learning post‑training.
On the DeepSeek‑V3 benchmark the NVL72 configuration scaled from 256 to 1,024 GPUs with roughly 97 % efficiency, cutting training time from 16 minutes to about four, while a 10,240‑GPU GB300 cluster demonstrated the largest scale in this round.
In its first MLPerf Endpoints entry, Vera Rubin delivered more than ten‑fold generational improvement in throughput per watt for agentic inference, measuring how much multi‑step reasoning work a system can perform per unit of power.
Why it matters for GPU and AI infrastructure
These results show that next‑generation GPU architectures combined with NVLink scale‑up and Quantum‑X800 InfiniBand scale‑out networking can keep thousands of accelerators working in concert. For cloud providers and enterprise buyers, that translates into faster model iteration, lower energy cost per token, and a clearer path to deploying agentic AI services at production scale.
- aigpu
- ai gpu
- ai gpu cloud
- aigpu dubai
- mlperf
- vera rubin
- nvidia
- gpu cloud
By AiGpu Editorial · Editorial rewrite based on public reporting (NVIDIA Blog)
← All articles