NVIDIA published Vera Rubin NVL72 preview results for MLPerf Inference v6.1 on September 16, covering the DeepSeek-R1 and Qwen3-VL models.
Reported performance
According to NVIDIA, throughput reached up to 3.7 times that of GB300 NVL72 on Qwen3-VL and up to 2.5 times on DeepSeek-R1. The company describes these as preview submissions.
NVIDIA also reported 99% scaling efficiency for a separate DeepSeek-R1 offline test that expanded GB300 NVL72 from one rack to four, using 288 GPUs.
Testing context
The results involve specific models, software stacks, and test scenarios. They do not establish an equivalent performance increase for every application.
NVIDIA separately discusses performance improvements recorded after the submission deadline and states that those additional results have not yet been verified by MLCommons.
What the announcement covers
- Vera Rubin NVL72 preview submissions.
- GB300 NVL72 scaling measurements.
- Software changes affecting inference throughput.
This report presents NVIDIA’s published results and qualifications; Guava has not independently tested the systems.
Source: NVIDIA benchmark announcement. Reporting checked September 16, 2026.
