Intel / blog

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

Source: blogs.nvidia.com
Author: Zhihan Jiang
Date: Published Sep 16, 2026

Context

NVIDIA reports that its Vera Rubin NVL72 system handled up to 3.7 times the throughput of GB300 NVL72 on Qwen3-VL and up to 2.5 times on DeepSeek-R1 in its MLPerf Inference v6.1 preview submissions. Throughput measures how much work a system completes over time, so these results concern serving capacity under the 's conditions.

NVIDIA attributes the results to full-stack codesign of hardware and software. It says the Vera Rubin submissions heavily used disaggregated serving, which separates processing the incoming prompt from generating the answer so each stage can be organized independently. The differing gains across models show why the workload matters when assessing a rack's capacity: the largest reported multiplier is not a universal speedup for every AI request.

Key takeaways · AI-distilled

NVIDIA's first MLPerf v6.1 submission for its next-gen Vera Rubin NVL72 system reported up to 3.7x the throughput of the current GB300 NVL72 on Qwen3-VL, and up to 2.5x higher throughput on DeepSeek-R1, per NVIDIA's own benchmark filing.
Scaling a DeepSeek-R1 workload from one GB300 NVL72 rack (72 GPUs) to four racks (288 GPUs) achieved 99% scaling efficiency in NVIDIA's MLPerf v6.1 submission, with throughput growing almost in direct proportion to the added hardware.
Software-only optimizations (lower KV-cache precision, kernel fusion, disaggregated prefill/decode serving) lifted GB300 NVL72's MLPerf score up to 1.6x between the v6.0 and v6.1 submissions, with no hardware change.
NVIDIA's sixth-generation NVLink and NVLink Switch interconnect claims 10x higher packet rates and 3x lower latency than off-the-shelf Ethernet, which NVIDIA says is what makes disaggregated serving and expert parallelism effective at rack scale.
On the SemiAnalysis AgentX benchmark, Vera Rubin NVL72 delivered 30x better performance than GB300 NVL72 in preview testing, a far bigger gap than on classic throughput benchmarks, reflecting the multi-step compute pattern of agentic workloads.

Terms in this piece · Glossary

inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Read the source blogs.nvidia.com