NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance
AI-assisted digest, checked against the original source before publishing. How we work
NVIDIA's Vera Rubin platform posts a clean sweep of MLPerf Training v6.1 and leads its first MLPerf Endpoints run, with scaling results up to 10,240 GPUs.
NVIDIA's Vera Rubin platform entered the MLPerf benchmark circuit with results released today, sweeping every MLPerf Training v6.1 benchmark and, in its first-ever MLPerf Endpoints submission, reporting more than 10x generational gains in throughput per watt for agentic AI — the measure of how much agentic work a system performs for each watt of power it draws.
On the training side, NVIDIA's platform was the only one with results on every benchmark this round, covering large language models, image generation, recommender systems and reinforcement learning post-training. In its debut, Vera Rubin NVL72 delivered up to 2x more performance per GPU than the highly optimized GB300 NVL72 system, which NVIDIA frames as an early read on generation-over-generation gains. Results were retrieved from MLCommons on November 16, 2026, and span eleven entries.
NVIDIA also put numbers behind running at scale. On the DeepSeek-V3 benchmark, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs at roughly 97% efficiency: quadrupling the GPU count cut time to train from 16 minutes to about four. The largest run this round used 10,240 GPUs on GB300 systems, and Microsoft Azure matched that count with 10,240 NVIDIA GPUs on GB200 systems. The company ties those results to NVLink scale-up networking that lets GPUs within a system behave as one, Quantum-X800 InfiniBand and Spectrum-X Ethernet linking systems across a data center, plus optimized software keeping thousands of GPUs in concert.
The stated use cases sit squarely with agents: NVIDIA points to AI writing code, running business workflows and answering customer questions, and to faster training shortening the path for drug-discovery models handling new lab data and fraud-detection models adapting to new threats — a theme this week's SC26 conference is highlighting around scientific research.
Looking ahead, NVIDIA says Vera Rubin performance will keep climbing as its CUDA software stack is continuously optimized, so infrastructure bought today gains value over time. What remains open is whether the 10x throughput-per-watt claim in the new Endpoints track holds across other vendors' agentic workloads, how closely follow-on results track the 2x per-GPU lead over GB300, and whether the near-97% scaling efficiency survives at the 10,240-GPU scale rather than the 1,024-GPU run reported here.
Key facts
- • MLPerf Training v6.1 results released November 16, 2026, listing eleven NVIDIA entries.
- • Microsoft Azure matched the scale run with 10,240 NVIDIA GB200 GPUs.
- • First MLPerf Endpoints submission reported 10x-plus throughput-per-watt gains for agentic AI.
- • NVIDIA was the only platform submitting results on every benchmark this round.
- • DeepSeek-V3 training fell from 16 minutes to about four by quadrupling GPUs.
Why it matters
Builders benchmarking training fleets now have a new per-GPU and per-watt yardstick: Vera Rubin NVL72 claims up to 2x GB300 NVL72 performance, and the Endpoints track makes power efficiency a comparable metric for agent workloads rather than raw speed alone. Expect procurement and capacity planning conversations to shift toward watts-per-agent-task, while the emphasis on software-driven, ongoing p
Spotted an error? Send a correction.
Get the day's AI news in one email
The Daily Epoch · 60-second read · unsubscribe anytime