Straight answer
● What it is NVIDIA's first MLPerf submission for Vera Rubin NVL72, in MLPerf Inference v6.1 - a round released on 16 September 2026.
● The numbers Up to 3.7x the throughput of GB300 NVL72 on Qwen3-VL, and up to 2.5x on DeepSeek-R1.
● The caveat that matters It is a PREVIEW submission - the first from that system - in the MLPerf Closed Division, and the figures are NVIDIA's own.
● Why it is here Vera Rubin is the silicon behind a large share of 2027 capacity commitments, and these are the only public numbers for it.
Two benchmarks were submitted for Vera Rubin NVL72: DeepSeek-R1, run on TensorRT-LLM, and Qwen3-VL, run on vLLM with NVIDIA Dynamo. The 3.7x figure is the Qwen3-VL result across offline, server and interactive scenarios; 2.5x is DeepSeek-R1. Both are per-rack comparisons against GB300 NVL72 rather than per-GPU, which is the right unit for a buyer costing a deployment and the wrong one for comparing chips.
The word doing the most work is preview. In MLPerf, a preview submission is for hardware not yet generally available, and it is not the same as a verified result on shipping silicon. NVIDIA's post also notes that its GPT-OSS-120B and DLRMv3 results were submitted after the deadline and are not yet verified. For scale context from the same round, the GB300 NVL72 submission ran 288 GPUs across four racks on DeepSeek-R1 offline and held 99% scaling efficiency from a single 72-GPU rack - a shipping-hardware number that is arguably more useful than the preview headline.
We are publishing this now because the round itself has had almost no coverage aimed at people choosing tools rather than buying racks, and because Vera Rubin keeps appearing in 2027 capacity announcements without anyone citing a number. If you are reading throughput claims in a vendor announcement over the next few months, these are the figures they are derived from, and the date on the round is 16 September rather than this week.
FAQ
Is this MLPerf Training or Inference?
Inference. MLPerf Inference v6.1, a round released on 16 September 2026. Several secondary write-ups have described it as a Training round, which it is not.
How much faster is Vera Rubin than GB300?
On NVIDIA's own figures, up to 3.7x the throughput on Qwen3-VL and up to 2.5x on DeepSeek-R1, compared per rack rather than per GPU.
What does preview submission mean?
It covers hardware that is not yet generally available. It is a legitimate MLPerf category, but it is not a verified result on shipping silicon and should not be read as one.
Which benchmarks did Vera Rubin submit?
DeepSeek-R1 on TensorRT-LLM and Qwen3-VL on vLLM with NVIDIA Dynamo, both in the Closed Division.