NVIDIA has unveiled the first performance results for Vera Rubin NVL72 in MLPerf Inference v6.1, delivering up to 3.7 times more performance than GB300 NVL72 in certain tests. The new platform also shows 99% scaling efficiency when moving from one to four GB300 NVL72 racks, while software optimizations boost performance by up to 1.6 times compared with the previous generation of the benchmark.
NVIDIA Vera Rubin NVL72 in 30 seconds
- Vera Rubin NVL72 reaches up to 3.7 times the performance of GB300 NVL72 on Qwen3-VL.
- On DeepSeek-R1, the reported advantage reaches up to 2.5 times depending on the scenario.
- GB300 NVL72 achieves 99% scaling efficiency with 288 GPUs across four racks.
- Software optimizations delivered up to 1.6 times more performance compared with MLPerf Inference v6.0.
- Vera Rubin’s results are still preview, while later figures for GPT-OSS-120B and DLRMv3 remain unverified.
The MLPerf Inference v6.1 results position Vera Rubin NVL72 as NVIDIA’s new platform for rack-scale inference workloads. Vera Rubin’s first participation in this benchmark uses the DeepSeek-R1 and Qwen3-VL models, two of the tests included in the new edition of the evaluation suite. MLCommons published the results on September 16, 2026, and notes that this edition also adds new tests for agentic AI and retrieval-augmented scenarios.
The comparison isn’t limited to GPU power. NVIDIA attributes part of the improvements to the combination of architecture, interconnect, and inference software. Specifically, the Vera Rubin tests use techniques such as disaggregated serving, which separates the prefill and decode phases, along with expert parallelism for mixture-of-experts architecture models.
Vera Rubin arrives at MLPerf with up to 3.7x more performance
In the Qwen3-VL tests, Vera Rubin NVL72 posts up to 3.7 times the performance of GB300 NVL72 across the offline, server, and interactive scenarios. NVIDIA used vLLM together with its open-source inference framework Dynamo for these tests.
With DeepSeek-R1, the reported difference reaches up to 2.5 times versus GB300 NVL72. In this case, NVIDIA used TensorRT-LLM, its library for optimizing language model inference.
The specific figures vary by workload type. In the results table published by NVIDIA, Qwen3-VL shows a 1.8x improvement in the offline scenario, 1.9x in server, and 3.7x in interactive. For DeepSeek-R1, the figures are 1.7x, 1.9x, and 2.5x, respectively.
The difference between scenarios matters because measuring only an infrastructure’s peak throughput doesn’t necessarily reflect how it will respond to real-world requests. The offline scenario aims to measure maximum aggregate throughput, while the interactive scenarios incorporate latency constraints.
MLCommons defines MLPerf Inference as a benchmark aimed at measuring inference system performance through reproducible tests with a shared methodology. The 6.1 edition has also added new tests for more complex workloads, including agentic inference.
Vera Rubin’s figures need a caveat: they are preview results. NVIDIA presented them as the platform’s first participation in MLPerf Inference 6.1, so they shouldn’t be interpreted as a definitive measurement across every possible configuration.
GB300 shows that adding GPUs can also scale nearly linearly
While Vera Rubin takes the spotlight in this new generation of results, NVIDIA also used MLPerf Inference v6.1 to show how GB300 NVL72 behaves as the number of racks increases.
The DeepSeek-R1 test compares a rack with 72 GPUs against a four-rack configuration with 288 GPUs. The result reaches 99% scaling efficiency in the offline scenario. The published table shows 679,740 performance units for one rack and 2,705,130 for four racks.
In practical terms, going from 72 to 288 GPUs quadruples the number of accelerators, and performance grows in almost the same proportion. Scaling efficiency measures precisely how much an increase in resources translates into an equivalent increase in performance.
That behavior depends on several factors. The performance of an installation with hundreds of GPUs doesn’t automatically increase just by adding more accelerators. Communication between GPUs, the network between racks, request distribution, and the software that coordinates the nodes can all introduce losses.
NVIDIA attributes part of this result to the NVL72 system and its NVLink interconnect. The sixth generation of NVLink and NVLink Switch provides the high-speed link within the 72-GPU domain, while other networking technologies handle communication between systems.
Software keeps changing the results after the hardware ships
Another data point from the 6.1 edition comes from optimizations made to GB300. NVIDIA says the performance achieved with Qwen3-VL improved by up to 1.6 times compared with its MLPerf Inference v6.0 results.
The improvements come from several changes to the software stack, including lower precision for the KV cache, new kernel fusions, optimizations to those components, and the use of disaggregated serving with vLLM and NVIDIA Dynamo.
This explains why NVIDIA presents AI platform performance as a combination of hardware and software. The same GPU generation can deliver different results as compilers, libraries, kernels, and workload management systems continue to evolve.
The company has also shown figures generated after the official MLPerf results submission. The published tables show 124,552 tokens per second on GPT-OSS-120B on B300 x8 in the offline scenario and 119,002 tokens per second in server. For DLRMv3 on B200 x8, the reported figures are 16,796 queries per second and 15,800 queries per second, respectively.
These figures are explicitly labeled as unverified results. NVIDIA states they correspond to optimizations made after the MLPerf Inference v6.1 submission deadline, so they don’t carry the same status as the results published and validated within the benchmark.
Software evolution also affects the economics of inference. If an infrastructure produces more tokens with the same hardware, an organization can spread the cost of servers, networking, and power across a larger volume of work. The end result depends on the specific workload, the model used, and the service conditions.
From traditional benchmarks to agentic AI
NVIDIA is also using agent-oriented tests to measure Vera Rubin. In preliminary SemiAnalysis AgentX evaluations cited by the company, Vera Rubin NVL72 reached 30 times the performance of GB300 NVL72. This figure comes from a preview test separate from MLPerf Inference and therefore shouldn’t be confused with official MLPerf results.
The shift in approach reflects the fact that AI agents can carry out several consecutive operations to complete a task. A traditional application might need just one response from the model, while an agent can reason, call tools, retrieve information, and run the model again multiple times.
MLCommons has started reflecting that shift in its own benchmarks. Inference v6.1 adds a test called Edge Agentic Inference and is preparing MLPerf Endpoints as an evolution for measuring inference workloads through APIs and scenarios closer to how production systems actually operate.
For NVIDIA, the goal is for the same infrastructure to be usable across different models and workloads, from training and inference to recommendation systems, reasoning, or video generation. That flexibility keeps the hardware from being tied to a single model as application needs change.
Vera Rubin NVL72 combines 72 Rubin GPUs, 36 Vera CPUs, and NVIDIA networking and processing components in a rack-scale architecture. The company is introducing it as a platform aimed at large-scale AI workloads and agentic systems.
The MLPerf Inference v6.1 results provide an initial public reference point for this new generation. The comparison with GB300 shows significant improvements in certain workloads, while the scaling tests demonstrate that the previous generation can also maintain high utilization when growing from one rack to four.
The open question is how these results will evolve once Vera Rubin moves past the preview stage and once the new platforms are measured against more models, configurations, and production workloads.
Frequently asked questions
How much faster is Vera Rubin NVL72 than GB300 NVL72?
In the MLPerf Inference v6.1 results published by NVIDIA, Vera Rubin NVL72 reaches up to 3.7 times the performance of GB300 NVL72 with Qwen3-VL and up to 2.5 times with DeepSeek-R1, depending on the scenario.
What does 99% scaling efficiency mean?
It means that when moving from one GB300 NVL72 rack with 72 GPUs to four racks with 288 GPUs, performance increases in almost the same proportion as the resources used in the offline test.
Are the Vera Rubin results already final?
No. NVIDIA presented the Vera Rubin NVL72 results as a preview within MLPerf Inference v6.1. The later figures for GPT-OSS-120B and DLRMv3 shown by NVIDIA are also not verified by MLCommons.
What role does software play in these improvements?
NVIDIA attributes part of the gains to software optimizations, such as changes to KV cache precision, kernel fusions, improved kernels, and disaggregated serving techniques. In its Qwen3-VL results, the company reports up to 1.6 times more performance compared with its MLPerf Inference v6.0 results.
