NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1

Summary

NVIDIA's Vera Rubin NVL72 made its debut with preview results in MLPerf Inference v6.1. The company claims 3.7x higher output on Qwen3-VL and 2.5x higher output on DeepSeek-R1 compared to the previous generation GB300 NVL72. Meanwhile, the GB300 demonstrated 99% scaling efficiency across 288 GPUs.

Why it matters

Inference economics—tokens per dollar, scaling efficiency, and software speed—now drive a significant portion of AI infrastructure spending. Consequently, a 3-4x generational jump in output directly reshapes data center capital expenditure planning. The 99% multi-rack scaling rate, if it holds up across various workloads, indicates that rack-scale NVLink connectivity has transitioned from a laboratory curiosity to a practical production standard. Additionally, the disaggregated prefill/decode presentation and the shift toward NVFP4 precision offer insights into where the next wave of inference stack design is heading.

Highlights

• Vera Rubin NVL72: 3.7x higher output on Qwen3-VL (vLLM + Dynamo) and 2.5x higher output on DeepSeek-R1 (TensorRT-LLM) compared to GB300 NVL72
• GB300 NVL72 achieved 99% scaling efficiency when scaling from 72 to 288 GPUs in the DeepSeek-R1 offline test
• Software optimizations alone between v6.0 and v6.1 yielded gains of up to 1.6x on Qwen3-VL; post-submission results for GPT-OSS-120B and DLRMv3 show additional, yet unverified, gains

Critical take

All headline figures are data points comparing NVIDIA's new chip against its older chip, as shared on NVIDIA's own blog. Therefore, there is no independent third-party hardware benchmark to contextualize the 3.7x claim. The most striking figure, a 30x increase in SemiAnalysis AgentX, is explicitly labeled as a 'preview test' and is not yet part of a standardized MLPerf run, making it the least verifiable number in the text.