Nvidia's Vera Rubin NVL72 Posts First MLPerf Inference Benchmark Results

Nvidia’s Vera Rubin NVL72 rack made its first appearance in the MLPerf Inference benchmark suite on September 16, 2026, with a preview submission to round v6.1. The company measured throughput against its own current-generation GB300 NVL72 rack, the platform it sells to cloud partners today, rather than against a rival’s hardware. On the Qwen3-VL workload, run through vLLM with Nvidia’s Dynamo inference framework, Vera Rubin NVL72 delivered higher throughput than GB300 NVL72 across offline, server and interactive scenarios. On DeepSeek-R1, run through the TensorRT-LLM library, the new rack again beat the older one. Nvidia credits the gains to Vera Rubin’s Tensor Cores and Transformer Engine, plus NVFP4 precision that shrinks the memory footprint of model weights, attention data and the key-value cache used during generation.

The same submission round carried a second result about the hardware already on sale: a GB300 NVL72 configuration scaled from a single rack up to four racks on the DeepSeek-R1 offline test, with throughput growing almost in step with the added hardware. That scaling result matters more than the headline throughput number for anyone buying racks today, because a system that does not scale cleanly erases the value of adding more of it. Nvidia also reported that software changes alone lifted GB300 NVL72’s Qwen3-VL score over the previous MLPerf round, through lower key-value cache precision, added kernel fusion and disaggregated serving that splits the prefill and decode stages of a request.

The numbers

Vera Rubin NVL72 posted up to 3.7x the throughput of GB300 NVL72 on Qwen3-VL and up to 2.5x on DeepSeek-R1, both measured in the same MLPerf Inference v6.1 preview round. The GB300 NVL72 scaling submission held 99% efficiency moving from a single rack to four, meaning throughput rose almost exactly in proportion to the added hardware rather than tapering off. Nvidia dates the software-only Qwen3-VL gain on GB300 NVL72 at up to 1.6x over the prior MLPerf round, separate from any new silicon.

None of these figures come from an independent buyer’s deployment. They are Nvidia’s own submission to MLCommons, checked against the benchmark’s rules but chosen and run by the vendor whose hardware is being measured. Preview status also means Vera Rubin NVL72 is not yet a shipping product with a public price, so a reader cannot convert the throughput gap into a cost-per-token figure today. What the release does establish is a baseline: the next generation of Nvidia’s inference hardware, once it ships, will be measured against these same numbers rather than against a competitor’s.