Original research brief
NVIDIA Vera Rubin NVL72 Performance in MLPerf Inference v6.1
NVIDIA Vera Rubin NVL72 preview results show throughput gains on DeepSeek-R1 and Qwen3-VL benchmarks, with scaling efficiency and software optimizations…
What this piece is grounded in
01
NVIDIA submitted Vera Rubin NVL72 preview results on DeepSeek-R1 and Qwen3-VL. (NVIDIA)
02
Vera Rubin NVL72 delivers up to 3.7x higher throughput than GB300 NVL72 on Qwen3-VL. (NVIDIA)
03
On DeepSeek-R1, throughput is up to 2.5x higher than GB300 NVL72. (NVIDIA)
04
NVIDIA's DeepSeek-R1 submission achieved 99% scaling efficiency in the offline scenario. (NVIDIA)
01 / FIELD NOTE
Vera Rubin NVL72 Benchmark Results
NVIDIA submitted Vera Rubin NVL72 preview results on two benchmarks in the MLPerf Inference v6.1 suite: DeepSeek-R1 and Qwen3-VL. According to the source, Vera Rubin NVL72 delivers up to 3.7x higher throughput than GB300 NVL72 on Qwen3-VL across offline, server and interactive scenarios. On DeepSeek-R1, throughput is up to 2.5x higher than GB300 NVL72. The source states these results reflect full-stack codesign across hardware and software, including enhanced Tensor Cores and Transformer Engine. The supplied feed entry is a concise summary; consult the canonical article for full context.
02 / FIELD NOTE
Scaling Efficiency and Software Gains
The source reports NVIDIA's DeepSeek-R1 submission scaled from a single GB300 NVL72 rack (72 GPUs) to four racks (288 GPUs), achieving 99% scaling efficiency in the offline scenario. It also notes that GB300 NVL72 performance on Qwen3-VL improved up to 1.6x over v6.0 results through software optimizations like lower KV cache precision and additional kernel fusion. The source mentions post-submission results on GPT-OSS-120B and DLRMv3 show further performance gains, though these are not yet verified by MLCommons.
03 / FIELD NOTE
What Changed and What Can Be Verified
The change is the debut of Vera Rubin NVL72 preview results in MLPerf Inference v6.1, showing specific throughput multipliers compared to the previous GB300 NVL72 platform on named benchmarks. A reader can verify the claimed throughput ratios of up to 3.7x on Qwen3-VL and up to 2.5x on DeepSeek-R1, as well as the 99% scaling efficiency for the GB300 NVL72 submission, by examining the official MLPerf Inference v6.1 results published by MLCommons. The source attributes these results to the use of vLLM with NVIDIA Dynamo and the TensorRT-LLM library.
04 / FIELD NOTE
Editorial Interpretation of Practical Implications
This editorial interpretation is based on source facts. The reported throughput increases suggest that, for data centers running the specified models, a Vera Rubin NVL72 rack could process more tokens per unit of time than its predecessor. The 99% scaling efficiency indicates that adding more GB300 NVL72 racks can yield near-linear throughput growth for the DeepSeek-R1 workload in an offline scenario, which matters for capacity planning. The source's mention of continuous software optimization implies that performance for these systems may improve further with subsequent software releases.
Questions readers ask
What benchmarks did NVIDIA use for the Vera Rubin NVL72 preview?
According to the source, NVIDIA submitted Vera Rubin NVL72 preview results on two of the most demanding benchmarks in the MLPerf Inference v6.1 suite: DeepSeek-R1 and Qwen3-VL.
What scaling efficiency was reported for GB300 NVL72?
The source states NVIDIA's DeepSeek-R1 submission scaled from a single GB300 NVL72 rack to four racks, achieving 99% scaling efficiency in the offline scenario.