Vera Rubin NVL72 MLPerf-Preview — nur NVIDIA-Zahlen

NVIDIA-Blog 2026-09-16. Preview MLPerf v6.1: bis ~3.7× / ~2.5× NVIDIA-claimed vs GB300. Kein absolutes #1 ohne MLCommons.

On 16 September 2026, NVIDIA published a first preview MLPerf Inference v6.1 submission for Vera Rubin NVL72. Headline figures are NVIDIA-claimed: up to ~3.7× throughput vs GB300 NVL72 on Qwen3-VL (vLLM + Dynamo) and up to ~2.5× on DeepSeek-R1 (TensorRT-LLM). MLCommons marks the configuration in a Preview category — early structured results, not proof of broad commercial availability or stable rental economics.

Editorial rule: do not crown absolute “#1” without MLCommons language and entry IDs. Quote ratios as NVIDIA-stated preview comparisons to like-sized GB300 NVL72, and note scenario variance (interactive gains can differ from offline/server). Power, TCO, and fleet reliability are out of scope for this draft.

MSP takeaway: useful for capacity planning conversations, but gate POCs on your own SLOs and software stack maturity — preview silicon ≠ production SLA.

FAQ

Offizielle MLCommons-Sieger?

NVIDIA-Preview-Submissions in MLPerf Inference v6.1. NVIDIA-claimed Ratios + Preview-Status; kein absolutes #1 ohne MLCommons-Framing.

Was MSPs weiter validieren sollten?

Software-Stack, Power/Cooling und Produktions-SLOs — Preview-Throughput ≠ Managed-Service-Garantie.

Sources

Draft only — do not publish without editorial review.