NVIDIA's Vera Rubin aces MLPerf debut with 10x agentic efficiency gain
NVIDIA's Vera Rubin platform swept every MLPerf Training v6.1 benchmark and logged more than 10x generational gains in throughput per watt on its first MLPerf Endpoints submission. The NVL72 system also scaled to 1,024 GPUs at 97% efficiency on DeepSeek-V3, with 20 partners backing the runs.
- Original source: blockchain.news.
- The source published it on Nov 17, 2026 at 2:24 AM.
- No second outlet has reported it yet. Treat it as a single-source claim.
NVIDIA's Vera Rubin platform cleaned house in the latest MLPerf round. Results released November 16, 2026 show the chipmaker taking all MLPerf Training v6.1 benchmarks and landing over 10x generational gains in throughput per watt in its first-ever MLPerf Endpoints submission. MLPerf Training tracks how fast systems train AI models to a target accuracy.
The Vera Rubin NVL72 rack delivered up to 2x more performance per GPU than the previous-generation GB300 NVL72. It posted results across every tested workload, spanning large language models, image generation, recommender systems and reinforcement learning post-training.
Scaling is the headline act. The NVL72 ran the DeepSeek-V3 benchmark from 256 up to 1,024 GPUs at 97% efficiency, and a job that took 16 minutes wrapped in four. NVIDIA also filed the round's largest submission, running the benchmark on 10,240 GB300 GPUs. The glue came from NVLink inside systems plus Quantum-X800 InfiniBand and Spectrum-X Ethernet across the data center.
In the Endpoints debut, Vera Rubin NVL72 served agentic AI workloads at more than 10x the throughput per watt of GB300 NVL72. On the Kimi K3 agentic workflow benchmark it hit 41,414.2 tokens per second system throughput with average interactivity of 50.4 tokens per second per user, burning 100.89 kW. GB300 NVL72 managed 2,210.4 tokens per second at 60 kW.
Twenty partners contributed to the training submissions, including Microsoft Azure, Google Cloud, Dell Technologies and Oracle. NVIDIA also pointed to independent results on SemiAnalysis AgentX and Artificial Analysis AgentPerf. The company says continued optimization of its CUDA-powered software stack will keep lifting Vera Rubin's performance for customers over time.
TopicsNVIDIA · Vera Rubin · GB300 NVL72 · MLPerf · DeepSeek-V3 · Microsoft Azure · Google Cloud · Dell Technologies
Source: blockchain.news
Written by AI from verified reporting.

