Ground Truth.
AI, checked against the source.

News · 2026-11-17

NVIDIA’s Rubin posts MLPerf training gains while serving result awaits review

NVIDIA’s Vera Rubin systems recorded approximately twice the training speed of matched-count GB300 systems on one MLPerf DeepSeek-V3 workload. A separate NVIDIA claim of more than tenfold serving throughput per watt comes from a provisional result with different operating conditions and is not yet an official reviewed comparison.

Key facts

Training benchmarks measure how long a system takes to reach a specified target on a defined workload. They compare entire machines and software stacks rather than revealing the speed of a single chip in isolation. The cited DeepSeek-V3 entries used the same accelerator count, numerical precision and named software release, making them more useful than comparing a larger machine with a smaller one. Their host processors and deployment configurations still differed.

At 256 accelerators, the corresponding entries took 16.124 minutes for Rubin and 31.404 minutes for GB300, also about a 1.95-fold difference. Scaling the Rubin run from 256 to 1,024 accelerators shortened its time by about 3.89 times for four times as many accelerators. That is the basis for NVIDIA’s roughly 97% strong-scaling efficiency: it compares Rubin with itself at two sizes, not Rubin with GB300. Near-linear scaling on this benchmark does not guarantee the same outcome for another model, network or cluster size.

Serving asks a different question: how quickly can an already trained model answer many users? NVIDIA’s announcement cites more than ten times throughput per watt for selected Kimi serving points. The submitted figures yield about 410.5 tokens per second per kilowatt for Rubin versus 36.0 for GB300, or roughly 11.4 times. This arithmetic uses the reported power at those operating points; a separate provisioned-power disclosure is not the same number.

The two points differ in important ways. Rubin used concurrency 1,024, compared with 48 for GB300, and the systems had 32 and 48 accelerators respectively. Rubin’s reported time to first token was about 4.67 seconds versus 2.36 seconds for GB300. A restaurant can serve more total meals per unit of energy while an individual diner still waits longer for the first course. The cited comparison concerns capacity at selected streaming speeds, not a guarantee that every part of a response appears faster.

MLCommons labeled the Rubin entry “Provisional / Review Pending” when checked, while the GB300 entry was official. Its result row does not supply a task-accuracy score. The numerical ratio is therefore a calculation from submitted system figures, not an independently established general claim about model quality or every serving workload. The Endpoints methodology provides the context for interpreting those operating points.

Power comparisons need the same care as time comparisons. A benchmark can report power at the operating point actually measured and separately disclose provisioned system power. The calculation above uses the point power: about 100.89 kilowatts for the selected Rubin run and 60 kilowatts for the GB300 point. Substituting a different disclosure without explaining the change would produce a different ratio. The submitted stacks also differ in client placement and network paths. Those details may seem peripheral, but serving performance is a property of the complete system, including how requests arrive and how the software batches them.

NVIDIA also announced Rubin HPC, a separate architecture it expects in 2028 and describes as having more than 50,000 CUDA cores and support for sparse workloads. That is a roadmap announcement; it does not include scientific application benchmarks. The useful takeaway is a strong, scoped training result, plus a serving claim that needs completed review and comparisons under conditions that matter to the intended application.

MLCommons calls the entry “Provisional / Review Pending,” a status that matters because publication on the dashboard is not the same as completing its review. The provisional status applies to the cited serving result, not the matched training records.


Primary source, verified: read the paper →

Key questions

How much faster was Rubin in the matched training comparison?

On the cited 1,024-accelerator DeepSeek-V3 training entries, Rubin took 4.14 minutes versus 8.03 minutes for GB300, about 1.94 times faster.

Is the 10-times serving result an official final MLPerf result?

No. The Rubin Endpoints entry was labeled provisional and review pending when the dossier checked it.

Does this prove every AI workload runs twice as fast?

No. The comparison is a system-level result on a particular benchmark, configuration and software stack.
Cite this

APA

Ground Truth. (2026, November 17). NVIDIA’s Rubin posts MLPerf training gains while serving result awaits review. Ground Truth. https://groundtruth.day/news/rubin-mlperf-training-and-provisional-serving-results.html

BibTeX

@misc{groundtruth:rubin-mlperf-training-and-provisional-serving-results,
  title  = {NVIDIA’s Rubin posts MLPerf training gains while serving result awaits review},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/rubin-mlperf-training-and-provisional-serving-results.html}
}

Topics: hardware · benchmarks · infrastructure

Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.