Lifestyle

NVIDIA Vera Rubin Debuts in MLPerf: Claims a Sweep of Training v6.1, With Over 10x Throughput per Watt on Kimi K3 (Provisional)

NVIDIA says its Vera Rubin platform posted leading results in MLPerf Training v6.1 and the debut MLPerf Endpoints, though the Vera Rubin Endpoints results are still pending MLCommons verification.

News date: About 4 min read

Share on LINE
Contents
  1. 1. What results did NVIDIA announce?
  2. 2. How reliable are the MLPerf Endpoints agentic inference results?
  3. 3. Which partners took part, and what does it mean for ordinary users?

On November 16, 2026, NVIDIA published the first MLPerf results for its Vera Rubin platform. MLPerf is a set of AI system performance benchmarks whose results are published by the MLCommons Association. The results cover training, post-training (further tuning after a model's initial training, such as reinforcement learning) and agentic inference. Agentic inference refers to an AI agent repeatedly reasoning, calling tools and checking its work to complete a task. The Training v6.1 results are published, while the Vera Rubin results in Endpoints are provisional and pending MLCommons verification.

What results did NVIDIA announce?

NVIDIA says it was the only platform in this round to submit results on every benchmark, spanning large language models, image generation, recommendation systems and reinforcement learning post-training. MLPerf Training measures how fast systems can train various AI models to a target accuracy.

NVIDIA says Vera Rubin NVL72 made its debut with up to 2x the per-GPU performance of its previous-generation GB300 NVL72 system. NVIDIA stresses these are early results and that it will keep optimizing through the CUDA software stack. On the DeepSeek-V3 benchmark, NVIDIA says Vera Rubin NVL72 scaled from 256 to 1,024 GPUs at about 97% efficiency, cutting training time from 16 minutes to about 4 minutes.

How reliable are the MLPerf Endpoints agentic inference results?

According to NVIDIA, MLPerf Endpoints tests systems with real agentic coding and enterprise workflows and checks that the agent still completes tasks correctly. Every step of AI computation produces tokens (the basic units AI uses to process and generate content). Throughput per watt measures how much agent work a system completes for each watt of power consumed. In the provisional Kimi K3 result, NVIDIA says Vera Rubin NVL72 delivered more than 10x the throughput per watt of GB300 NVL72 at the same response speed.

NVIDIA's footnote states that the GB300 results have been verified by MLCommons, while the Vera Rubin provisional results await verification and may change with peer review. The 10x figure should therefore be treated as preliminary, vendor-published data.

MLPerf Endpoints Kimi K3 run data from NVIDIA's footnote (comparison baseline: at least 50 tokens/sec per user)
ItemVera Rubin NVL72GB300 NVL72
Kimi K3 accelerators3248
Concurrent queries1,02448
System throughput (tokens/sec)41,414.22,210.4
Operating power100.89 kW60 kW
Verification statusProvisional, pending MLCommons verificationVerified by MLCommons

Which partners took part, and what does it mean for ordinary users?

NVIDIA says 20 organizations submitted results to MLPerf Training, including ASUS, Azure, Cisco, CoreWeave, Dell Technologies, Google Cloud, HPE, Oracle, Supermicro and the University of Florida. NVIDIA also says it ran the round's largest test, with 10,240 GPUs on GB300 systems. NVIDIA says Microsoft Azure matched that scale, also using 10,240 GPUs, on GB200 systems.

NVIDIA's argument is that faster training gets smarter agents into users' hands sooner, while more tokens per watt in inference means power-constrained data centers can serve more agents. NVIDIA also says Vera Rubin leads on independent benchmarks such as SemiAnalysis AgentX and Artificial Analysis AgentPerf.

Frequently asked questions

Can I buy or use Vera Rubin yet?

NVIDIA's article only publishes benchmark results and does not state availability timing, pricing or regional supply, so this cannot be determined from the source.

What does "sweep" mean?

NVIDIA says its platform was the only one in this round to submit results on every MLPerf Training v6.1 benchmark, and says it swept all of them. This is NVIDIA's own characterization.

Is the 10x throughput per watt a final result?

No. NVIDIA notes the Kimi K3 Vera Rubin result is provisional, pending MLCommons verification, and may change with peer review.

Will the results change later?

NVIDIA says Vera Rubin performance will keep improving through continued optimization of the CUDA software stack, and calls the current results just the beginning.

Browse the latest news in this topic

Found this guide helpful? Share it with someone who needs it

Share on LINE

Sources

Compiled from the references and verification dates listed above; fares, schedules, product features, and terms can change, so please confirm the latest details on the relevant official website before travelling, booking, or purchasing.

Lifestyle