Vera Rubin NVL72 Posts Up to 3.7x Over GB300 in Its First MLPerf Inference v6.1 Submission

NVIDIA published the first preview submission results for Vera Rubin NVL72 alongside the MLPerf Inference v6.1 round on September 16, 2026. Vera Rubin NVL72 delivers up to 3.7x higher throughput than GB300 NVL72 on Qwen3-VL across the offline, server and interactive scenarios, up to 2.5x higher on DeepSeek-R1, and 30x better performance than GB300 NVL72 on SemiAnalysis AgentX agent reasoning in preview testing. The same announcement carries GB300 NVL72 results as well: 99% scaling efficiency in the offline scenario when DeepSeek-R1 scales from one rack of 72 GPUs to four racks of 288 GPUs, and up to 1.6x on Qwen3-VL from v6.0 to v6.1 through software alone. ASAP separates which hardware and which condition each of these numbers belongs to.

The numbers in this release belong to two different systems

The scorecard splits between Vera Rubin NVL72 and GB300 NVL72, with the headline multiples attached to the former and the scaling and software-gain figures to the latter. The numbers on Vera Rubin NVL72 are up to 3.7x on Qwen3-VL and up to 2.5x on DeepSeek-R1, both measured against GB300 NVL72. The Qwen3-VL submission ran on vLLM with NVIDIA Dynamo, and the DeepSeek-R1 submission ran on TensorRT-LLM.

The GB300 NVL72 figures are a different kind of claim. Scaling DeepSeek-R1 from one rack of 72 GPUs to four racks of 288 GPUs produced 99% scaling efficiency in the offline scenario, and rack-scale testing on the WAN 2.2 text-to-video benchmark produced 0.65 720p videos per second at 5.7 seconds per video, which NVIDIA states is 9x higher throughput and 7.5x lower latency than a single node. The up to 1.6x on Qwen3-VL from v6.0 to v6.1 is also a GB300 NVL72 result.

NVIDIA lists submission entries 6.1-0106, 6.1-0074 and 6.1-0073. It also states that software optimization continued past the v6.1 submission deadline, and that further gains on GPT-OSS-120B and DLRMv3 are post-submission results not yet verified by MLCommons. That distinction has to survive every time these figures are quoted.

Four levers are given as the reason for the speedup

The four levers NVIDIA names as the source of the speedup are numeric precision, serving structure, parallelism strategy and interconnect. The first is NVFP4 precision, a 4-bit floating-point format applied to inference so that more tokens move through the same memory and bandwidth. The second is disaggregated serving, which places prefill and decode on separate resources to exploit the different characteristics of the two phases.

The third is expert parallelism for mixture-of-experts models. Both DeepSeek-R1 and Qwen3-VL belong to families that use mixture-of-experts structures, so spreading experts across many GPUs and waking only the routed ones pays off at rack scale. The fourth is sixth-generation NVLink, which NVIDIA describes as offering 10x higher packet rates and 3x lower latency than off-the-shelf Ethernet.

All four point in one direction. Today's inference bottleneck is the cost of moving tokens and KV cache inside a rack rather than the arithmetic capability of a single GPU. Multiples such as 3.7x and 2.5x are not the product of a chip generation swap on its own; they carry the effect of lowering precision, splitting phases and rerouting communication. This matters for procurement because the software share of that gain applies retroactively to hardware already installed.

30x and 1.6x are opposite kinds of number

The largest multiple in the release, 30x, is not an MLPerf submission result. NVIDIA describes it as 30x better performance than GB300 NVL72 on SemiAnalysis AgentX AI agent reasoning in preview testing. That places it outside the benchmark whose rules MLCommons sets and verifies, and outside the set of results whose comparison conditions and workload definitions a third party can reproduce. Quoting 30x requires carrying the preview-testing qualifier with it.

The 1.6x on the other side is the more useful number in practice. It is the difference between running Qwen3-VL on v6.0 and on v6.1 on the same GB300 NVL72 hardware, where the only thing that changed is software. Throughput on an already purchased and installed rack rose by as much as 1.6x across a single benchmark cycle.

For teams computing inference cost these two numbers support opposite conclusions. The 30x is weak grounds for waiting on the next generation, while the 1.6x is solid grounds for expecting that the cost of existing capacity still has room to fall. A plan that lowers cost per token only through hardware replacement discards the software share for free. Conversely, a baseline measured on a stack that has not been upgraded to v6.1 inflates the apparent need to replace anything.

The word offline attached to 99% scaling efficiency

The 99% scaling efficiency figure carries offline scenario as its condition. MLPerf Inference measures separate scenarios, and offline submits requests in bulk and measures total throughput. The server and interactive scenarios impose response-latency constraints and behave differently, and the target NVIDIA reported 99% against is offline.

That condition feeds straight into service design. Offline throughput growing almost linearly out to four racks is good news for batch inference. Bulk document embedding, overnight classification passes, data labeling and synthetic data generation all have loose latency budgets and gain in proportion to racks added. Conversational services and real-time agents cannot assume the same curve.

That the Qwen3-VL submission spans offline, server and interactive together matters in this contrast. The phrase up to 3.7x points at the most favorable of the three scenarios as an upper bound, and the per-scenario values differ. The right order of work is to decide which scenario your own latency budget resembles and then look up the figure for that scenario. Copying the upper bound alone puts capacity planning off course.

What the words preview submission mean for a purchase plan

Vera Rubin NVL72's result here is a preview submission rather than a closed-division one, and NVIDIA itself calls it a first MLPerf Inference preview submission with preview results. The preview category in MLPerf applies to systems that are not yet generally available, and unlike the closed category it presumes room for the numbers to shift in the shipping configuration.

Anyone evaluating a purchase should read that word alongside delivery timing. The 3.7x and 2.5x are indicators of where the next rack generation is headed, not verified performance of a product available to buy. The GB300 NVL72 figures in the same release, by contrast, come from a system already in operation. Teams planning capacity for late 2026 and the first half of 2027 should baseline on the latter and use the former as material for roadmap and negotiation.

The post-submission label belongs to the same discipline. The additional gains on GPT-OSS-120B and DLRMv3 come from optimization after the deadline and have not been verified by MLCommons. This is precisely where a benchmark regime earns its keep. Mixing a number a third party confirmed under shared rules and accuracy targets with a number the vendor measured itself, inside a single sentence, destroys comparability. When these move into a procurement document, verification status belongs in its own column.

A new edge benchmark marks where the next contest sits

The new entry in v6.1 is the Edge-Agentic benchmark, and NVIDIA submitted Jetson AGX Thor results on it using TensorRT Edge-LLM with Qwen3.6-27B. Separate from the rack-scale contest, agent workloads measured on a device have now entered the benchmark suite.

The reason a new entry matters is that benchmarks follow the industry's attention. A workload usually appears in MLPerf around the time it starts appearing in procurement documents. Creating an entry that measures an agent calling tools and advancing through multiple steps on the edge signals that device performance is to be judged by multi-step task completion rather than single-shot inference latency.

For manufacturing, robotics and automotive software teams this entry becomes a concrete reference point, because a public measurement now exists for running a 27-billion-parameter-class model as an agent on an edge module. What this announcement contains, though, is the fact of the submission and its configuration; no throughput or power figures accompany it. Any team evaluating on-device agents should read the raw numbers for this entry directly from the v6.1 results table MLCommons publishes.

Source: NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut (NVIDIA, September 16, 2026)