NVIDIA's Vera Rubin NVL72 has made its MLPerf Inference benchmark debut, with preview submissions for DeepSeek-R1 and Qwen3-VL in the v6.1 results released September 16. MLCommons separately confirms that Rubin and Vera Rubin NVL72 appear in the preview category, rather than among the round's available systems.

NVIDIA reports up to 3.7 times higher throughput than GB300 NVL72 on Qwen3-VL across the tested offline, server and interactive scenarios. It reports up to 2.5 times higher throughput on DeepSeek-R1. These are NVIDIA's comparisons of specific benchmark configurations, not a guarantee of equivalent gains for every application.

Hardware and serving software

The Qwen3-VL submission uses vLLM with NVIDIA Dynamo, while the DeepSeek-R1 submission uses TensorRT-LLM. NVIDIA says its Rubin submissions make extensive use of disaggregated serving, separating prompt processing from token generation, alongside expert parallelism for mixture-of-experts models.

That matters for infrastructure buyers: the results measure a hardware-and-software system, not an accelerator operating independently of its serving stack. NVIDIA also distinguishes post-submission software measurements from results verified by MLCommons.

Benchmarks follow changing workloads

The wider release introduces end-to-end retrieval-augmented generation and edge agentic inference tests. MLCommons says the RAG benchmark measures both document ingestion and query answering through a pipeline involving embeddings, retrieval, re-ranking and language models. The edge test targets multi-turn workloads such as agentic coding.

The release provides a new reference point for comparing inference systems. Deployment decisions still require matching model, latency target, system size and availability category; headline throughput alone does not establish production cost or application quality.