MLCommons released MLPerf Training v6.1 on November 16, including NVIDIA's Vera Rubin systems. Results
The big change
- In the DeepSeek-V3 671B benchmark run, Rubin took 4.14 minutes versus GB300's 8.03 minutes, using 1,024 GPUs each.
- The benchmark starts from a prepared checkpoint and measures progress toward a defined quality target.
- A rack's processors and interconnections form a system; the chip name alone does not describe it.
What the matched comparison shows
In the Closed DeepSeek-V3 671B benchmark, submission 6.1-0116 used 1,024 Rubin GPUs and scored 4.14 minutes. The 1,024-GPU GB300 submission, 6.1-0030, scored 8.03 minutes, approximately 1.94 times as long. Each used 256 nodes with four GPUs per node. Both used NeMo 26.09 and CUDA 13.4.1 software, with FP8 (eight-bit floating-point) linear operations. Their host CPUs and networks differed, so this is a system comparison rather than an isolated GPU test. Explorer
Scale changes the ranking. A 10,240-GPU GB300 submission, 6.1-0029, finished in 1.38 minutes. The selected Rubin entry is Available on-premise; the selected GB300 entries are Available cloud. Submissions
What those minutes measure
MLPerf Training measures time to a specified quality threshold. This task uses C4 text and a validation-loss target of 3.60, where lower loss means better performance on the evaluation. The reference procedure begins from an already trained, prepared DeepSeek checkpoint. Four minutes therefore does not mean building the entire model from scratch. Reference
The Closed division constrains the model and training procedure to make comparisons meaningful while permitting implementation differences. Repeated runs and scoring rules address variation; a headline score should not be treated as a hand-picked stopwatch reading. Nor does meeting this target establish performance on a company's own documents or application. Rules
The Available category means systems can be purchased or rented. It differs from Preview, which carries a later availability requirement. That classification does not itself establish a particular buyer's delivery date. Categories
Why the rest of the system matters
NVIDIA describes its NVL72 platform as a rack combining 72 Rubin GPUs and 36 Vera CPUs. NVLink connects processors within the rack; network fabrics connect larger deployments. Those are different parts of the communication path, and a rack specification does not tell a buyer how many processors a particular job used. Platform
This distinction matters when assessing a proposed cluster. Ask which rack configuration, host processors and interconnections the offer includes, rather than treating every deployment bearing the same platform name as equivalent. The product page describes an architecture and advertises availability; it does not demonstrate the customer's workload or settle its commercial terms.
The practical next step is a comparison using the intended job and offered configuration, alongside price, power and delivery information. BIG CHANGE inspected the public records; it did not rerun the benchmark or test a customer deployment.



