Latest news
HardwareNVIDIANVIDIA Vera Rubin

NVIDIA Reports MLPerf Leadership for Vera Rubin

NVIDIA says Vera Rubin submitted results in every MLPerf Training v6.1 test and stood out in MLPerf Endpoints agent benchmarks. The Kimi K3 comparison is still awaiting validation.

An architectural-scale model seen behind glass beside server racksAI image
Representative image, generated with AI.Image: 3dsınıfı / FCA AI

In brief

  1. NVIDIA said Vera Rubin was the only platform to submit results in every test in the MLPerf Training v6.1 round.
  2. According to the company, Vera Rubin NVL72 achieved approximately 97% efficiency as it scaled from 256 to 1,024 GPUs in the DeepSeek-V3 test.
  3. NVIDIA reported more than 10 times the throughput per watt in the Kimi K3 test compared with GB300 NVL72; the result is preliminary and awaiting validation.
  4. The company said Vera Rubin NVL72 delivered up to twice the performance per GPU of GB300 NVL72.

NVIDIA published Vera Rubin platform results for MLPerf Training v6.1 and MLPerf Endpoints on 16 November 2026. The company said Vera Rubin stood out in training and post-training workloads, as well as AI agent inference tests. According to NVIDIA, the platform was the only system to submit results in every MLPerf Training test in this round and also achieved leading results across all MLPerf Endpoints agent benchmarks.

What results were shared for the training tests?

MLPerf Training measures how quickly systems can train different AI models to a target accuracy level. The tests NVIDIA shared cover large language models, image generation, recommendation systems, and reinforcement learning post-training. The company says Vera Rubin NVL72 achieved up to twice the performance per GPU of an optimized NVIDIA GB300 NVL72 system. This is the manufacturer’s claim based on initial results; the announcement does not show that the same gain was achieved in every workload.

NVIDIA said the Vera Rubin NVL72 system scaled from 256 to 1,024 GPUs in the DeepSeek-V3 test, achieving approximately 97% efficiency. In the company’s measurement, increasing the GPU count fourfold reduced training time from 16 minutes to around four minutes. NVIDIA attributes the systems’ ability to work together at large training scales to its NVLink scale-up interconnect, along with Quantum-X800 InfiniBand and Spectrum-X Ethernet networks; it says the results were achieved with optimized software.

Is the more-than-10x result in the agent test definitive?

MLPerf Endpoints evaluates agent tasks such as coding and enterprise workflows, and whether those tasks are completed correctly. NVIDIA reported more than 10 times the throughput per watt for Vera Rubin NVL72 compared with GB300 NVL72 on the Kimi K3 model. The company’s comparison measures throughput per watt at a shared interaction threshold of at least 50 tokens per second per user.

This is not a final, independently validated result. NVIDIA says the Kimi K3 result reported for Vera Rubin is awaiting MLCommons validation and may change during peer review. The performance ratio should therefore be treated as the company’s preliminary result until validation is complete. Metrics such as test conditions and response speed should also be considered when comparing different systems.

Impact on architecture and visualization workflows

The announcement does not test architectural software, 3D rendering, or specific visualization tools. The results therefore cannot be interpreted as evidence that Vera Rubin directly reduces render times or speeds up a particular design application. Its potential indirect significance relates to more efficient training and agent workloads running behind the AI services used by design teams.

For firms considering adding AI agents to their workflows, the announcement is relevant as an indication of progress in infrastructure performance; however, it does not say that specific services use Vera Rubin. NVIDIA’s announcement also does not disclose the product’s price, release date, or compatibility with architectural software. When evaluating hardware investments or workflow changes, it is therefore important to consider not only the manufacturer’s benchmark results but also their validation status and the conditions of the intended use.

Sources

1 source
NB
NVIDIA Blogblogs.nvidia.com/blog/mlperf-training-endpoints-vera-rubin
Summary

Source texts are not republished; short quotes are marked, everything else is our own summary and commentary.

3dsınıfı’s take
3dEditor’s assessment

The Vera Rubin results show that, alongside training performance, agent throughput per unit of energy is becoming increasingly important in AI infrastructure. For architecture firms, this is not a promise about render times or design software performance today; rather, it is a development that could have implications for the infrastructure behind the AI services they use.

For teams in Turkey, actual availability, total cost, licensing, and compatibility with the software in use will be key factors in purchasing decisions. Since the Kimi K3 result has not yet been validated, it would be unwise to treat the reported figures as definitive performance expectations. A more cautious approach is to monitor the announcement until commercial terms and independent validation become clear.

Frequently asked questions

What result did NVIDIA report for Vera Rubin in MLPerf Training v6.1?

NVIDIA said Vera Rubin was the only platform to submit results in every MLPerf Training test in this round. The tests cover large language models, image generation, recommendation systems, and reinforcement learning post-training.

Has Vera Rubin’s more-than-10x result in the Kimi K3 test been validated?

No. NVIDIA says the Vera Rubin result is preliminary and awaiting MLCommons validation and peer review.

Does Vera Rubin improve architectural visualization performance?

The tests in the announcement do not measure architectural software or 3D rendering performance. These results cannot establish that a specific architectural workflow will be faster.

Comments and the forum are in Turkish.Join the discussion
+

Related news