Vera Rubin NVL72 Debuts in MLPerf Training v6.1
MLCommons published the MLPerf Training v6.1 results with 153 submissions from 24 organizations. The NVIDIA Vera Rubin NVL72 platform appeared in the training benchmark for the first time.

Vera Rubin NVL72 appeared for the first time in the MLPerf Training v6.1 results published by MLCommons on November 16, 2026. The round included 153 submissions from 24 organizations. In its interpretation of the results, NVIDIA reported the fastest time on six of the seven benchmark tasks.
MLPerf is a suite of benchmarks used to compare AI infrastructure on precisely defined tasks. The current training results therefore provide a public comparison point for NVIDIA’s new platform, but they are not a general measurement of performance or energy efficiency across all production deployments.
Vera Rubin NVL72 in Its First Training Round
NVIDIA says its systems achieved the fastest times on six of the seven MLPerf Training v6.1 tasks. For the DeepSeek-V3 671B task, the company reported the best time for a GB300 NVL72 system in a cluster with 10,240 GPUs.
The results are not exclusively NVIDIA results. MLCommons recorded submissions from 24 organizations in this round, and the benchmark includes multiple system and infrastructure vendors. Comparing individual entries requires examining the specific task, configuration, and conditions of each run.
Verified Endpoint Result for Kimi K3
Alongside the training round, MLCommons lists NVIDIA results in MLPerf Endpoints v1.0, which focuses on model operation under defined endpoint conditions. For the Kimi K3 model, a verified run from November 4 is available for Vera Rubin NVL72.
This run achieved 9,940 system tokens per second at a P90 TTFT of 1,286 milliseconds. TTFT, or time to first token, expresses the latency before the system begins generating a response; the P90 value represents the 90th percentile of measured cases. However, the figures cannot be directly compared with all other results without further context, since endpoint tests depend on the model, mode, configured latency, concurrency, and power consumption.
The Energy Claim Remains Preliminary
NVIDIA also claims more than a tenfold generational improvement in throughput per watt for agentic AI compared with GB300 at specified interactivity. However, the Vera Rubin result used for this claim comes from November 9 and has a “Provisional” status in the MLCommons database, meaning it is a preliminary result awaiting peer review.
This status does not mean the result is invalid, but that its verification process has not yet been completed. Published and verified endpoint runs should therefore be distinguished from preliminary entries, especially when interpreting energy efficiency. The initial Vera Rubin NVL72 results alone do not confirm tenfold better efficiency in typical operation.
What to Watch Next
It will be important to see whether MLCommons completes verification of the preliminary Vera Rubin endpoint result and whether its values change. Further comparable results against systems from AMD, Google, and other vendors under identical models and operating conditions will also be relevant. The availability of Vera Rubin NVL72 systems from cloud providers and the energy characteristics of real-world deployments remain open practical questions.
Sources
- MLCommons — MLPerf Training v6.1 Results – Confirms the publication of the results on November 16, 2026, 153 submissions, and participation by 24 organizations.
- MLCommons Results — Benchmark Results – Confirms the current MLPerf Training v6.1 and MLPerf Endpoints v1.0 rounds and the number of published submissions.
- MLCommons Results — Vera Rubin NVL72, Kimi K3 – Documents the MLCommons-verified Vera Rubin NVL72 result in MLPerf Endpoints v1.0.
- MLCommons Results — Preliminary Vera Rubin NVL72, Kimi K3 – Confirms that the result used in the energy claim has “Provisional” status.
- NVIDIA Technical Blog — Vera Rubin NVL72 MLPerf Training v6.1 – Presents NVIDIA’s interpretation: Vera Rubin had the fastest time on six of seven tasks, and GB300 on the DeepSeek-V3 671B task.
Verified and updated: November 17, 2026, 06:24



