NVIDIA’s September 16 MLPerf Inference v6.1 post describes preview results for Vera Rubin NVL72 and submissions involving GB300 NVL72. The company reports throughput and scaling gains on specified benchmarks, while separately discussing software improvements made after the submission period.
Read the dated primary source ↓How to read a result
A benchmark measures a defined system under defined conditions. Offline throughput, interactive latency and task quality are not interchangeable. Preview hardware results and post-submission figures also need to remain distinct from the formally submitted comparisons.
What buyers should compare
Inspect the model, precision, software stack and scenario behind any headline number. Test a representative workload and include power, capacity and operational constraints. A strong result can justify further evaluation without establishing that the same proportional gain will appear in every application.
What this means in practice
Inference benchmarks are meaningful when model, precision, workload and quality targets are visible. A headline throughput result should not be directly translated into every application's response time or cost. Compare the benchmark conditions with the actual service before drawing a deployment conclusion.
Questions readers ask
Can one inference benchmark identify the best hardware for every task?
No. Results depend on workload and test conditions. Compare relevant quality, latency, throughput and operating constraints for the intended application.
Follow the evidence.
Source-based coverage of the dated announcement below, with Promptwire analysis. Company-reported claims are attributed to their source. The update date reflects added context and answers; it does not change the original event date.
Public sources support the factual background. Analysis and illustrative exercises are identified in the text. How we assess AI · Editorial standards
Spot a factual error or an outdated detail?



