Physical AI News (Japan)

Physical AI, robotics and compute — reported from Japan.

日本語
· 日本語で読む

NVIDIA reports MLPerf inference performance for Vera Rubin NVL72, up to 3.7x vs. GB300

NVIDIA reported preview performance for Vera Rubin NVL72 in MLPerf Inference v6.1. For Qwen3-VL, it said the system delivered up to 3.7x the throughput of GB300 NVL72. Meanwhile, GB300 NVL72 recorded 99% scaling efficiency in a 288-GPU configuration.

Key facts

  • NVIDIA said Vera Rubin NVL72 delivered up to 3.7x the throughput of GB300 NVL72 for Qwen3-VL.
  • NVIDIA said Vera Rubin NVL72 delivered up to 2.5x the throughput of GB300 NVL72 for DeepSeek-R1.
  • GB300 NVL72 scaled from 72 to 288 GPUs in an offline DeepSeek-R1 test, recording 99% scaling efficiency.
  • NVIDIA said Vera Rubin NVL72 delivered 30x the performance of GB300 NVL72 in a preview test of SemiAnalysis AgentX.
  • NVIDIA said GB300 NVL72 performance on Qwen3-VL improved by up to 1.6x versus v6.0 in MLPerf Inference v6.1.

What happened

NVIDIA released preview results for Vera Rubin NVL72 in MLPerf Inference v6.1. It said the system recorded up to 3.7x the throughput of GB300 NVL72 for Qwen3-VL across offline, server and interactive scenarios, and up to 2.5x for DeepSeek-R1. In an offline DeepSeek-R1 test, GB300 NVL72 scaled from one 72-GPU rack to four racks with 288 GPUs, recording 99% scaling efficiency.

Why it matters

According to NVIDIA's published results, GB300 NVL72 achieved 99% scaling efficiency when expanded from 72 to 288 GPUs, while Vera Rubin's advantage varies by model and benchmark. Organizations choosing AI inference infrastructure need to compare not only rack performance but also the target model, inference software, efficiency as GPUs are added and the status of results verification.

Analysis

The announcement centers not only on peak performance from a new-generation rack but also on comparisons that include inference configurations and software updates. The up-to-3.7x figure for Vera Rubin NVL72 applies to Qwen3-VL, while the up-to-2.5x figure applies to DeepSeek-R1; neither should be treated as a general performance difference under identical models and conditions. Both are also preview results for Vera Rubin and should be viewed as benchmark figures reported by NVIDIA. Meanwhile, the 99% figure for GB300 NVL72 reflects efficiency when scaling from 72 to 288 GPUs in an offline DeepSeek-R1 test. It is not a general rule that processing volume will increase at the same rate as the number of GPUs; it is a result for this model, configuration and test conditions. It is also important that Vera Rubin's figures combine hardware with software such as NVFP4, disaggregated prefill and decode, vLLM and NVIDIA Dynamo. Evaluations of inference infrastructure need to distinguish rack-level performance from the efficiency of the overall operation, including models and processing methods. A related article from this newspaper, Lambda connects GB300 NVL72 racks with NVIDIA CPO switches, with more than 10,000 GPUs, covered the connection of a rack cluster with more than 10,000 GPUs. Its scale and focus differ from the 288-GPU test here. The 99% figure cannot be used to estimate the efficiency of a large-scale network, but the comparison shows that connections within and between racks are another area to examine alongside the expansion of computing resources. Another related article, Nikkei publishes analysis of confidence in NVIDIA after ‘Mr. Huang’, is confirmed by the information provided only to have been published with that headline; its argument cannot be linked to this performance assessment. Published benchmark figures should be considered separately from independent verification and real-world performance. At an industry level, NVIDIA's comparison highlights an approach that optimizes not just chip performance but also in-rack connections, inference software, numerical precision and task partitioning as a whole. The company's explanation that software optimization alone raised Qwen3-VL performance by up to 1.6x versus v6.0 is evidence that infrastructure economics are not determined solely by hardware specifications at the time of adoption. However, results for GPT-OSS-120B and DLRMv3 submitted after the deadline have not been verified by MLCommons, and the 30x AgentX result is also from a preview test. Points to watch include formal verification results, benchmark conditions and output quality, and whether similar performance can be achieved under actual workloads.

Details

The Vera Rubin NVL72 Qwen3-VL test used vLLM and NVIDIA Dynamo, while the DeepSeek-R1 test used TensorRT-LLM. NVIDIA described faster inference processing with Tensor Cores and Transformer Engine, and reduced memory use for model weights, Attention and KV cache through NVFP4. It also used disaggregated serving, which separates prefill and decode, and expert parallelism for Mixture-of-Experts layers. NVIDIA said its sixth-generation NVLink and NVLink Switch deliver 10x the packet rate and one-third the latency of commercial Ethernet. According to NVIDIA, Vera Rubin NVL72 delivered 30x the performance of GB300 NVL72 in a preview test of SemiAnalysis AgentX. In a WAN 2.2 text-to-video generation test, GB300 NVL72 recorded 0.65 videos per second and 5.7 seconds per video, with 9x the throughput and one-seventh the latency of a single node. NVIDIA said GB300 NVL72 performance on Qwen3-VL improved by up to 1.6x versus v6.0, and noted that results for GPT-OSS-120B and DLRMv3 submitted after the deadline had not yet been verified by MLCommons. Nebius also submitted preview results for Vera Rubin NVL72, but no individual figures were provided.

Sources

Translated from the Japanese original: /news/nvidia-vera-rubin-nvl72-no-mlperf-suiron-seinou-wo-kouhyou-g-c60864. The Japanese version is authoritative if they differ.