NVIDIA today said its new Vera Rubin AI platform achieved leading results in the MLPerf Training v6.1 and Endpoints benchmarks, claiming up to twice the performance per GPU compared with its GB300 NVL72 system and more than ten times the throughput per watt for agentic AI workloads. The Vera Rubin Endpoints result is provisional and has not yet been officially verified by the MLCommons Association.
NVIDIA said the Vera Rubin platform posted results on every MLPerf Training v6.1 benchmark, making it the only entrant with full participation. In the company's comparison, Vera Rubin NVL72 delivered up to twice the performance per GPU compared with the already optimized GB300 NVL72 system.
MLPerf Training measures how quickly a system can train AI models to a target accuracy, covering large language models, image generation, recommender systems and reinforcement learning. The companion MLPerf Endpoints benchmark evaluates systems on agentic tasks such as coding and enterprise workflows, checking that agents complete their tasks correctly.
On the Endpoints tests, NVIDIA reported that Vera Rubin NVL72 achieved more than ten times the throughput per watt of a GB300 NVL72 system at the same response speed, using the Kimi K3 open model for agentic workflows. The Vera Rubin result is provisional and not yet verified by the MLCommons Association.
NVIDIA also reported scaling efficiency: on the DeepSeek-V3 model, increasing GPUs from 256 to 1,024 cut training time from 16 minutes to about four minutes, representing roughly 97% efficiency. NVIDIA additionally ran the benchmark on 10,240 GB300 GPUs, the largest scale in the round, matched by Microsoft Azure's 10,240-GPU GB200 system.
For training at scale, NVIDIA said its NVLink scale-up networking connects GPUs within systems, while Quantum-X800 InfiniBand and Spectrum-X Ethernet link those systems across the data center. Combined with optimized software, this keeps thousands of GPUs working in concert. The company said Vera Rubin's performance will continue to improve through optimization of its CUDA-based software stack, allowing customers' infrastructure to deliver more value over time.
Twenty organizations submitted MLPerf Training results using NVIDIA platforms, including ASUS, Azure, CoreWeave, Dell Technologies, Fujitsu, Google Cloud, HPE, Oracle and Supermicro. NVIDIA also noted that Vera Rubin leads on third-party benchmarks such as SemiAnalysis AgentX and Artificial Analysis AgentPerf.
Analysis
Why This Matters
- Faster AI training and more efficient inference can accelerate deployment of AI agents in coding, customer service, fraud detection, and science, potentially lowering costs for companies and data centers.
- NVIDIA's claims of up to 2x per-GPU performance and 10x throughput-per-watt improvements indicate significant efficiency gains, which could affect energy consumption and operational costs for AI infrastructure.
- The debut of a new platform often triggers competition and further innovation in the AI hardware market.
Background
NVIDIA is a leading designer of graphics processing units (GPUs) and AI accelerator hardware. Its previous platform, GB300, was highly optimized, and the new Vera Rubin platform is being positioned as a major upgrade. MLPerf, run by the MLCommons Association, is an industry-standard benchmark suite for measuring AI system performance. The results reported are from the company's own submissions and some are provisional pending verification.
What to Watch
- The final, verified MLPerf Endpoints results for Vera Rubin, which may confirm or adjust the provisional figures.
- How competitors (e.g., AMD, Intel, startups) respond with their own platforms in future MLPerf rounds.
- Real-world adoption: whether customers and cloud providers deploy Vera Rubin at scale and report independent performance numbers.
Checked
12 claims in this story checked against 1 source: 11 supported.
Show each claim
- SupportedNVIDIA said Vera Rubin led MLPerf Training v6.1 and Endpoints in its debut.NVIDIA Blog: “In its MLPerf benchmarks debut, with results released today, the NVIDIA Vera Rubin platform delivers leading performance across training, post-training and agentic inference.”
- SupportedVera Rubin delivered up to twice the performance per GPU of the GB300 NVL72 system.NVIDIA Blog: “In its debut, Vera Rubin NVL72 delivered up to 2x more performance per GPU than the highly optimized NVIDIA GB300 NVL72 system”
- Partly supportedVera Rubin delivered more than ten times the throughput per watt of the previous-generation system.NVIDIA Blog: “in a provisional result for Kimi K3, a large open model built for agentic workflows, Vera Rubin NVL72 systems deliver more than 10x the throughput per watt of NVIDIA GB300 NVL72 at the same response speed. ... results are subject to change pending peer review.”
- SupportedThe Vera Rubin Endpoints result is provisional and has not yet been verified by the MLCommons Association.NVIDIA Blog: “For the provisional Vera Rubin results: pending MLCommons Association verification, results are subject to change pending peer review.”
- SupportedNVIDIA was the only entrant with results on every MLPerf Training v6.1 benchmark.NVIDIA Blog: “The NVIDIA platform was the only one with results on every benchmark this round”
- SupportedOn the Kimi K3 Endpoints test, Vera Rubin delivered more than ten times GB300’s throughput per watt at the same response speed.NVIDIA Blog: “in a provisional result for Kimi K3, a large open model built for agentic workflows, Vera Rubin NVL72 systems deliver more than 10x the throughput per watt of NVIDIA GB300 NVL72 at the same response speed.”
- SupportedOn DeepSeek-V3, scaling from 256 to 1,024 GPUs cut training time from 16 minutes to about four minutes, for roughly 97% efficiency.NVIDIA Blog: “Vera Rubin NVL72 scaled from 256 to 1,024 GPUs with about 97% efficiency: quadrupling the GPUs cut time to train from 16 minutes to about four.”
- SupportedNVIDIA ran the benchmark on 10,240 GB300 GPUs, the largest scale in the round.NVIDIA Blog: “NVIDIA also ran the benchmark on 10,240 GPUs with GB300 systems, representing the largest scale in this round.”
- SupportedMicrosoft Azure matched that scale with a 10,240-GPU GB200 system.NVIDIA Blog: “Microsoft Azure matched that scale with 10,240 NVIDIA GPUs with GB200 systems.”
- SupportedNVLink connects GPUs within systems, while Quantum-X800 InfiniBand and Spectrum-X Ethernet link systems across the data center.NVIDIA Blog: “NVIDIA NVLink scale-up networking connects the GPUs within each system so they act as one, while NVIDIA Quantum-X800 InfiniBand and Spectrum-X Ethernet scale-out networking link those systems across the data center.”
- SupportedTwenty organizations submitted MLPerf Training results using NVIDIA platforms, including the organizations named in the story.NVIDIA Blog: “NVIDIA partners participated extensively in MLPerf Training, with submissions from 20 organizations: ASUS, Azure, Cisco, CoreWeave, Crusoe, Dell Technologies, Firmus, Fujitsu, Giga Computing, Google Cloud, HPE, Inventec, Krai, Lambda, Nebius, Oracle, Quanta Cloud Technology (QCT), Supermicro, TTA and the University of Florida.”
- SupportedNVIDIA said Vera Rubin leads on SemiAnalysis AgentX and Artificial Analysis AgentPerf.NVIDIA Blog: “Vera Rubin also delivers leading performance on independent benchmarks including SemiAnalysis AgentX and Artificial Analysis AgentPerf”