NVIDIA's next-generation Vera Rubin platform has made its first appearance in MLPerf Training results. In a company post published November 16, NVIDIA says the NVL72 system delivered up to twice the performance per GPU of its GB300 NVL72 system in selected training workloads. The results are notable because MLPerf is a public benchmark suite run by MLCommons, allowing system submissions and configurations to be examined instead of relying only on a vendor's own demonstration.
What the benchmark measures
MLPerf Training measures the time a system takes to train a model to a specified target level of quality. Its workloads span more than one type of computing task, including large language models and image generation. A result is meaningful only with its workload, system configuration and completion criteria attached. It is not the same as measuring how quickly a chatbot answers a user, nor does a training result automatically predict the performance of every model an organization might build.
NVIDIA says it was the only company with results submitted across every benchmark in the v6.1 training round. That breadth gives readers a full set of entries from one platform, but it does not mean every entry is a direct contest between identical systems from different suppliers. The MLCommons results page and the benchmark's rules are the right places to inspect which systems submitted each workload and how the tested configurations differ.
The scale-up test behind the 97 percent figure
One of the clearest figures in NVIDIA's post concerns DeepSeek-V3 training. The company reports scaling the Vera Rubin system from 256 to 1,024 GPUs while maintaining about 97 percent efficiency. In its account, increasing the GPU count fourfold reduced training time from roughly 16 minutes to about four. MLCommons' supplemental discussion gives the more specific times as 16.12 and 4.14 minutes, which represents close to a 3.9-times speedup.
That kind of scaling test asks whether extra accelerators do useful work together instead of waiting on communication or software bottlenecks. For large training runs, the connections inside a server and between servers can matter as much as the speed of a single chip. Still, strong scaling on one named workload is not a promise that every customer can reproduce the same result with a different data pipeline or model.
How to interpret the performance claim
NVIDIA also reported submissions at 10,240 GPUs using GB300 systems, while Microsoft Azure matched that scale with GB200 hardware. That comparison illustrates the growing size of benchmark runs, but those two demonstrations used different GPU generations. For companies budgeting a data center, elapsed training time is only one input: power, cooling, hardware availability, system cost and software requirements affect how quickly a deployment pays off.
The new Vera Rubin entries are an early benchmark snapshot, not a full review of shipping systems under everyday production loads. MLPerf provides a more structured comparison than a vendor slide deck, yet careful readers should look at each row, note which configuration is being compared and avoid turning a workload-specific score into a general claim that one GPU is always twice as fast. The useful story is the combination of first published results and a large-scale efficiency demonstration, with the detailed entries left available for technical scrutiny.
Why it matters
The MLPerf debut gives the market a concrete reference point for Vera Rubin, and the scaling result is more informative than an unqualified peak-performance number because it shows what happened as the GPU count increased. Even so, the strongest headline needs two qualifiers: it compares selected benchmark workloads, and the result comes from NVIDIA's own description of its submission. MLCommons publishes the underlying entries, which are the place to check the exact configurations. Buyers should pair those scores with cost, power and production measurements before translating benchmark speed into an infrastructure decision.
Sources and further reading
This story is written in its own words. These links identify the reporting and announcements used to verify the details above.
Facts checked Nov 21, 2026.




