Vera Rubin's MLPerf debut: 2x training, 10x agent efficiency

MLPerf Training v6.1 results released November 16, 2026 gave NVIDIA's Vera Rubin platform its benchmark debut: the fastest time to train on six of seven benchmarks, up to 2x performance per GPU over GB300 NVL72, and a provisional 10x throughput-per-watt lead on agentic inference. Here is what the numbers say, what is still provisional, and what it means if you buy GPU capacity.
- Up to 2x per GPU on training — In MLPerf Training v6.1 closed division, Vera Rubin NVL72 trained DeepSeek-V3 671B in 4.14 minutes on 1,024 GPUs versus 8.03 minutes for GB300 NVL72 at the same scale; other benchmarks showed 1.4x to 1.6x per-GPU gains.
- Agentic efficiency lead is provisional — On MLPerf Endpoints v1.0, Vera Rubin NVL72 showed more than 10x the throughput per kilowatt of GB300 NVL72 on Kimi K3 at a shared 50 tokens/sec/user interactivity floor. The Vera Rubin results are provisional, pending MLCommons verification.
- Scaling held at 97% efficiency — Vera Rubin NVL72 kept above 97% overall efficiency scaling DeepSeek-V3 training from 256 to 1,024 GPUs, enabled by an NVLink 6 spine at 3.6 TB/s per GPU plus Quantum-X800 InfiniBand or Spectrum-X Ethernet.
- Generally available through cloud partners — NVIDIA says Vera Rubin NVL72 systems are generally available through cloud infrastructure partners, and 20 partner organizations submitted MLPerf Training results on NVIDIA platforms this round.
What MLPerf Training v6.1 actually measured
Related guide: Google AI Mode Hits 1 Billion Monthly Users: What Search's Agent Actually Does
MLPerf Training measures wall-clock time to train a model to a fixed quality target - not peak FLOPS, not a marketing slide. The v6.1 suite spans seven workloads: DeepSeek-V3 671B and gpt-oss 20B mixture-of-experts pretraining, Qwen3.5 397B GRPO reinforcement-learning post-training, Llama 3.1 8B pretraining, Llama 2 70B LoRA fine-tuning, FLUX.1 image generation, and the DLRMv4 recommender. The closed division requires the same reference models, which makes it the apples-to-apples comparison.
NVIDIA's technical blog publishes the per-benchmark ratios against GB300 NVL72 at matched GPU counts: 2x on DeepSeek-V3, 1.6x on gpt-oss 20B and Llama 3.1 8B, 1.5x on Llama 2 70B LoRA and FLUX.1, and 1.4x on Qwen3.5 GRPO and DLRMv4. The pattern is informative: the biggest gain lands on the largest, most communication-heavy model, where the new NVLink 6 fabric - a copper spine delivering 3.6 TB/s per GPU - does the most work. NVIDIA platforms were also the only ones with results on every benchmark, and 20 organizations submitted NVIDIA-based results, from Azure and Google Cloud to CoreWeave, Lambda, Nebius and Oracle. Breadth of submission matters: it shows the software stack works across the ecosystem, not just inside NVIDIA's lab.
The headline run: DeepSeek-V3 in 4.14 minutes
Related guide: NVIDIA RTX Spark desktops: local AI on Windows
The standout result is the 671-billion-parameter DeepSeek-V3 benchmark. Vera Rubin NVL72 trained it to target quality in 4.14 minutes on 1,024 GPUs across 16 racks; GB300 NVL72 needed 8.03 minutes on the identical 1,024-GPU configuration. Halving time to train roughly doubles how many experiments a team can run per quarter, and experiment throughput is what moves model quality.
Just as important is how the system scaled. Vera Rubin kept above 97% overall efficiency going from 256 to 1,024 GPUs - quadrupling the hardware cut training time from about 16 minutes to about four. Adding GPUs only helps if the fabric keeps them working together, and this is the number that says the rack-scale design holds. At the extreme end, GB300 produced the round's fastest absolute DeepSeek-V3 time, 1.38 minutes on 10,240 GPUs across 160 racks, with Microsoft Azure matching that scale on GB200 systems. The message NVIDIA is sending: buy Vera Rubin for per-GPU efficiency, trust the deployed GB300 fleet for massive scale today.
One more detail worth noting for buyers: NVIDIA also ran unverified reruns with a newer software stack that further cut training times on Llama 3.1 8B, gpt-oss 20B and DLRMv4. The company's pitch is that the CUDA stack keeps improving after the hardware ships - a real effect, but the published, verified numbers are the ones above, not the reruns.
Agentic inference: 10x per watt, with an asterisk
Related guide: IBM and Microsoft Launch Imagine Enterprise Agent Platform
The second half of the story is MLPerf Endpoints v1.0, a newer benchmark that tests real agentic coding and enterprise workflows through API endpoints and checks that the agents still complete tasks correctly. Agents are brutal on inference systems: context grows across turns - the benchmark's median trajectory grows from about 30,000 tokens at the first response to about 200,000 by response 300 - and every tool call generates more tokens. Because AI data centers are power-limited, the metric that matters is throughput per watt at an acceptable response speed.
On the Kimi K3 agentic benchmark, provisional Vera Rubin NVL72 results show more than 10x the throughput per kilowatt of GB300 NVL72 at a shared minimum of 50 tokens/sec per user: 41,414 tokens/sec system throughput at 100.89 kW serving 1,024 concurrent queries, versus 2,210 tokens/sec at 60 kW for GB300 at its qualifying operating point. Across DeepSeek-V4.1-Flash and Qwen3.6-35B-A3B the ratios are 9x and 5x per kilowatt, with 3x to 5x interactivity gains at matched concurrency.
The asterisk is real and NVIDIA states it plainly: the Vera Rubin Endpoints results are provisional, pending MLCommons verification, and subject to change after peer review; the GB300 results are verified. The comparison also selects each platform's best qualifying operating point, which is standard MLPerf practice but means the 10x is a best-case efficiency ratio, not a blanket statement about every deployment. If agentic serving economics drive your roadmap - and after this year's agent adoption they usually do - this is the number to watch when verification completes. It rhymes with the efficiency argument we covered in the Q3 FY27 earnings preview: the constraint on AI growth is power, and tokens per watt is the currency.
The SC26 context: Rubin's roadmap beyond the GPU
Related guide: Gemini 4 Neon: Benchmarks, Pricing & Antigravity Guide
The MLPerf debut sits inside a broader SC26 push. NVIDIA announced the Vera CPU rack-scale platform for traditional scientific computing: 192 CPUs with 88 custom Olympus cores each, about 300TB of memory, Quantum-X800 InfiniBand, and NVIDIA-measured results of 1.9x performance and 2.4x performance per watt versus x86 on OpenFOAM's motorBike simulation. The German Climate Computing Center (DKRZ) has selected Vera CPUs for its HLRE-5 climate supercomputer, and Vera CPU racks are expected from Dell, HPE, Lenovo and Supermicro in early 2027. NVIDIA also announced Rubin HPC, a simulation-focused GPU with more than 50,000 CUDA cores aimed at agent-driven science workflows - expected in 2028.
Read together, the announcements say NVIDIA is building for a world where agents don't just generate tokens but run simulations, schedule experiments and drive scientific loops. That is a roadmap claim, not shipped product; the shipped, benchmarked thing today is Vera Rubin NVL72 for AI training and inference, which NVIDIA says is generally available through cloud infrastructure partners. For the leadership context around all of this - Jensen Huang's November 12 resignation and Colette Kress's expanded role - see our earlier CEO transition analysis.
What this means if you buy compute
Related guide: Nvidia's quarter: $96.2B revenue, a $108B next-quarter guide, and a record $150B buyback top-up
Three practical takeaways. First, if you rent GPU capacity, ask your cloud provider when Vera Rubin instances arrive and at what price per training-hour - a 1.4x to 2x per-GPU gain changes the math even at a premium hourly rate. Second, if you serve agents at scale, wait for the MLCommons verification of the Endpoints results before rewriting your capacity plan, but start measuring your own tokens-per-watt or tokens-per-dollar now, because that is the axis on which the next generation competes. Third, keep the vendor caveat attached to every number here: these are NVIDIA's submissions and NVIDIA's blog posts about an audited benchmark - the audit is real, the framing is still marketing. The verified training sweep is solid evidence of a generational step; the 10x agentic figure is a provisional best-case until peer review says otherwise.
Frequently asked questions
Up to 2x per GPU on training
In MLPerf Training v6.1 closed division, Vera Rubin NVL72 trained DeepSeek-V3 671B in 4.14 minutes on 1,024 GPUs versus 8.03 minutes for GB300 NVL72 at the same scale; other benchmarks showed 1.4x to 1.6x per-GPU gains.
Agentic efficiency lead is provisional
On MLPerf Endpoints v1.0, Vera Rubin NVL72 showed more than 10x the throughput per kilowatt of GB300 NVL72 on Kimi K3 at a shared 50 tokens/sec/user interactivity floor. The Vera Rubin results are provisional, pending MLCommons verification.
Scaling held at 97% efficiency
Vera Rubin NVL72 kept above 97% overall efficiency scaling DeepSeek-V3 training from 256 to 1,024 GPUs, enabled by an NVLink 6 spine at 3.6 TB/s per GPU plus Quantum-X800 InfiniBand or Spectrum-X Ethernet.
Generally available through cloud partners
NVIDIA says Vera Rubin NVL72 systems are generally available through cloud infrastructure partners, and 20 partner organizations submitted MLPerf Training results on NVIDIA platforms this round.
- NVIDIA Blog - NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts (November 16, 2026)
- NVIDIA Technical Blog - Vera Rubin NVL72 Achieves up to 2x Blackwell Ultra Performance in MLPerf Training v6.1
- NVIDIA Technical Blog - Vera Rubin NVL72 Leads Agentic Inference Performance on MLPerf Endpoints v1.0
- MLCommons - MLPerf Training benchmark suite and rules
- NVIDIA Newsroom - NVIDIA Expands Vera CPU Platform to Accelerate Scientific Discovery (November 16, 2026)
- NVIDIA Blog - NVIDIA Announces Rubin HPC (November 16, 2026)
Prices and benchmark figures are standard published rates as of November 18, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.