In short
In short: MLPerf is a set of speed tests for AI systems, run by the engineering consortium MLCommons. The first tests, for training, launched in May 2018. MLPerf Inference measures how fast a system processes inputs and produces results with an already trained model. In the closed division everyone uses the same reference model, so the comparison is apples to apples.
Picture a car magazine that drives every car on the same track, on the same day, with the same driver. Each brand's ad says something different. The track shows one thing. MLPerf is the track.
The three words that decide it
Closed or open division. In the closed one the model is fixed and the comparison is fair. In the open one you can swap the model or retrain it, so the numbers don't compare directly with the others.
Available, preview or research. Available means you can buy or rent it today. Preview means the system must be submittable as available in the next round. Research means experimental or internal hardware. Before you believe the big number in the headline, check which box it came from.
An example from September 2026: in MLPerf Inference v6.1 NVIDIA submitted Vera Rubin NVL72 for the first time, as a preview result. By NVIDIA's numbers the system delivers up to 3.7 times the throughput of the previous generation, GB300 NVL72. The result is in the closed division, but it's for a machine that still has to become available.
November 2026 added two new rounds. In MLPerf Training v6.1 NVIDIA swept every benchmark and was the only one with a result on each of them. The debut of Vera Rubin NVL72 delivers up to 2x more performance per GPU than GB300 NVL72. On the DeepSeek-V3 benchmark the system went from 256 to 1,024 GPUs and cut training from 16 minutes to about four, at about 97 percent scaling efficiency, by NVIDIA's numbers.
In its first MLPerf Endpoints submission, the new round for agentic tasks, NVIDIA delivers more than 10x the throughput per watt of GB300 NVL72 at the same response speed. Those numbers are provisional: they await MLCommons verification and are subject to change.
The record in the keynote and the record in the table aren't always the same record.
Primary sources
MLCommons: MLPerf Inference DatacenterMLCommons: about the consortiumNVIDIA: Vera Rubin NVL72 in MLPerf Inference v6.1NVIDIA: MLPerf Training v6.1 and Endpoints v1.0The visual is generated code-art, no third-party images.