Every generation of Nvidia hardware arrives with a benchmark story, and Vera Rubin’s is the most sweeping yet. On November 16, the company released its first MLPerf results for the platform: a clean sweep of every benchmark in the MLPerf Training v6.1 round, plus a first submission to MLPerf Endpoints, the newer suite that measures how efficiently systems run multi-step AI agents. The headline numbers are large — up to twice the per-GPU training performance of the GB300 systems Vera Rubin succeeds, and more than ten times the throughput per watt on an agentic workload. The second of those two numbers, though, carries a caveat Nvidia itself prints in the footnotes: it is provisional, and MLCommons has not verified it yet.

A clean sweep in training#

MLPerf Training measures how fast a system can train a set of models to a target accuracy, and the v6.1 round covers large language models, image generation, recommender systems and reinforcement-learning post-training. Nvidia says its platform was the only one with results on every benchmark in the round. In its debut, Vera Rubin NVL72 delivered up to twice the performance per GPU of the GB300 NVL72 system — a comparison Nvidia frames as early-stage, noting performance should keep climbing as its CUDA software stack is tuned.

The scaling figure is the one that matters most to buyers sizing clusters. On the DeepSeek-V3 benchmark, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs at about 97% efficiency: quadrupling the hardware cut training time from 16 minutes to about four. Nvidia also ran the same benchmark on 10,240 GB300 GPUs, which it calls the largest scale in this round, and Microsoft Azure matched that scale with 10,240 GB200 GPUs. The point of the exercise is that adding GPUs only speeds up training if the systems work together efficiently — Nvidia credits its NVLink scale-up networking inside each system, plus Quantum-X800 InfiniBand and Spectrum-X Ethernet linking systems across a data center.

A close-up of a GPU chip package with gold memory stacks and glowing green circuitry
AI-generated illustration: the chip at the center of the benchmark claims.

The 10x agentic claim — and its asterisk#

The more dramatic number comes from MLPerf Endpoints, which tests systems on real agentic coding and enterprise workflows and checks that the agents still complete their tasks correctly. Because data centers are limited by power, the metric that counts is throughput per watt: how much agentic work a system delivers for each watt it draws. On a provisional Kimi K3 result — Kimi K3 is a large open model built for agentic workflows — Vera Rubin NVL72 delivered more than ten times the throughput per watt of GB300 NVL72 at the same response speed, according to Nvidia.

The raw figures show why the asterisk matters. The Vera Rubin run used 32 accelerators at 1,024 concurrent queries, posting 41,414.2 tokens per second at 100.89 kW; the GB300 run used 48 accelerators at 48 concurrent queries, posting 2,210.4 tokens per second at 60 kW. Nvidia’s footnote is explicit: the GB300 figures are verified by the MLCommons Association, while the Vera Rubin figures are provisional, pending verification and subject to change on peer review. A 10x ratio built from one specific configuration pairing is not the same as a general claim that Vera Rubin answers agents ten times better — and until MLCommons signs off, it should be read as a submission, not a confirmed result.

Glowing light columns of rising height above a dark server floor, an abstract image of climbing benchmark performance
AI-generated illustration: the generational gains, in abstract.

Why the round matters beyond Nvidia#

The Training round was not a one-vendor exercise. Twenty organizations submitted results on Nvidia’s platform, including Azure, CoreWeave, Dell Technologies, Google Cloud, HPE, Lambda, Oracle, Supermicro and the University of Florida — which Nvidia presents as evidence that its programmable platform lets partners optimize training performance independently. The company also cites leading results on independent benchmarks beyond MLPerf, including SemiAnalysis AgentX and Artificial Analysis AgentPerf, though it does not detail how those were configured or verified.

The timing adds edge to the numbers. Four days earlier, AMD launched the Instinct MI450X and claimed 30% faster inference than Vera Rubin, setting up a direct generational contest in which both vendors are now trading benchmark claims. And Vera Rubin is already moving into the market through other channels: Google Cloud’s A5X instances bring it to the cloud with general availability in November, and SpaceX signed a $40 billion lease for Vera Rubin capacity in October. What Nvidia has not said is when the hardware itself ships broadly, or at what price — the release discloses neither.

What to watch#

Three checkpoints will settle how much of this debut holds. First, MLCommons verification of the provisional Endpoints results — if the 10x figure survives peer review, it becomes the number that matters for anyone sizing power budgets around agentic workloads; if it is revised down, the Training sweep still stands on its own. Second, follow-on MLPerf rounds, which will show whether the 2x per-GPU training gain grows as the software stack matures, as Nvidia promises, or was a first-run peak. And third, the commercial facts the release omits: pricing, availability, and how much of this performance reaches customers who are not running Nvidia’s largest configurations. Benchmarks set the terms of the argument; the market decides whether they were the right terms.

Sources: Nvidia Blog, “NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance”; BlockWest, “Nvidia sweeps MLPerf training with Vera Rubin platform debut”; Silicon Report, “Nvidia debuts Vera Rubin with leading MLPerf benchmark results”; alfw.ai, “Nvidia Vera Rubin MLPerf 10x Is Provisional, Not Verified”.