On November 16, 2026, NVIDIA announced the first MLPerf benchmark results for its next-generation AI platform, Vera Rubin NVL72. In training DeepSeek-V3 671B on 1,024 GPUs, Vera Rubin finished in 4.14 minutes, compared with 8.03 minutes for a GB300 configuration using the same number of GPUs. NVIDIA also presented a comparison in which Vera Rubin's inference throughput per unit of power was more than 10 times higher, but the Vera Rubin figures are provisional results still awaiting verification by MLCommons.
The two results measure AI platform performance from different angles. MLPerf Training measures the time it takes for a model to reach a specified level of quality. The MLPerf Endpoints comparison, by contrast, sets a lower limit on per-user generation speed and then evaluates the total system throughput and operating power that meet that condition. Training time and inference power efficiency are separate metrics, and the results need to be read alongside the number of GPUs available and the response requirements of the service.
Training time roughly halved on the same 1,024 GPUs; the fastest record is 1.38 minutes on GB300
The MLPerf Training v6.1 results published by MLCommons include measurements for Vera Rubin and GB300 submitted by NVIDIA. In training DeepSeek-V3 671B, Vera Rubin on 1,024 GPUs recorded 4.14 minutes and GB300 recorded 8.03 minutes, so at the same GPU count the training time was about half. The fastest record of 1.38 minutes, however, came from a large-scale configuration of 10,240 GB300 GPUs. NVIDIA's technical blog and MLCommons' supplemental materials also show how performance varies with GPU count.
| Model | GPU generation | GPUs | Training time | Purpose of comparison | Source status |
|---|---|---|---|---|---|
| DeepSeek-V3 671B | Vera Rubin | 256 | 16.12 min | Scaling within one generation | NVIDIA's explanation in MLCommons supplemental materials |
| DeepSeek-V3 671B | Vera Rubin | 1,024 | 4.14 min | Generational comparison at the same GPU count | NVIDIA submission |
| DeepSeek-V3 671B | GB300 | 1,024 | 8.03 min | Generational comparison at the same GPU count | NVIDIA submission |
| DeepSeek-V3 671B | GB300 | 10,240 | 1.38 min | Fastest time at large scale | Explained in NVIDIA's technical blog |
All figures are times for DeepSeek-V3 671B to reach the specified quality in MLPerf Training v6.1.
The GB300 system that recorded the 1.38-minute time spans 160 racks and uses 10 times as many GPUs as the 1,024-GPU configuration. Comparing generations at the same GPU count and shortening training time by expanding the scale of the facility rest on different premises. Both 1,024-GPU configurations used FP8 and NeMo Framework Release 26.09, but Vera Rubin was registered as a configuration available on-premises, while GB300 was registered as available in the cloud.
A performance gap also appeared in post-training, the stage that follows initial model training. In the newly added reinforcement learning task "Qwen3.5 397B GRPO," Vera Rubin with 1,024 GPUs recorded 10.44 minutes and GB300 recorded 13.78 minutes. However, Vera Rubin used NeMo RL with TensorRT-LLM, while GB300 used NeMo RL with vLLM, so this is a comparison of whole platforms, including software, rather than of the GPUs alone.
Post-training with reinforcement learning repeats a cycle of generating answers, evaluating them, and updating model weights. According to NVIDIA's technical blog, even on the same Vera Rubin hardware, the time was 10.94 minutes with the vLLM backend and 10.44 minutes with TensorRT-LLM. Because the choice of software changes training time, real-world performance cannot be judged from GPU generation alone.
Scaling efficiency of about 97% as GPUs are added, with networking becoming critical
Adding GPUs lets the computation be divided among more devices, but it also increases the time spent exchanging data between them. In the MLCommons supplemental materials, NVIDIA explains that increasing Vera Rubin fourfold, from 256 to 1,024 GPUs, cut DeepSeek-V3 training time from 16.12 minutes to 4.14 minutes. That is roughly a 3.9x speedup, for a scaling efficiency, meaning the share of the increase in GPU count that translates into performance gains, of about 97%. This figure is based on NVIDIA's own explanation in the MLCommons supplemental materials.
NVIDIA cites high-speed communication within and between racks as the technology behind this result. Vera Rubin NVL72 houses 72 GPUs and 36 CPUs in a single rack, connecting the GPUs with copper-cabled NVLink 6. The published interconnect bandwidth is 3.6 TB per second per GPU and 260 TB per second for the whole rack. Multiple racks are connected using ConnectX-9 SuperNIC, with either Quantum-X800 InfiniBand or Spectrum-X Ethernet.
Mixture-of-experts (MoE) models such as DeepSeek-V3 route each input to several expert networks. This requires exchanging data across the regions handled by different GPUs, so communication speed and efficiency, not just the raw compute of a single GPU, affect training time. NVIDIA explains that mechanisms for selecting communication paths dynamically and for controlling network congestion support this kind of processing.
In training, how quickly large-scale computation can be completed determines the time needed for model development. In inference services, by contrast, what matters is how many requests can be processed with limited power while maintaining the response speed users expect. NVIDIA's claim of more than 10x performance per unit of power evaluates the latter.
Inference power efficiency more than 10x, but comparison conditions and verification status differ
NVIDIA's announcement says that, in the MLPerf Endpoints v1.0 results for Kimi K3 published on November 9, Vera Rubin's throughput per unit of power exceeded GB300's by more than 10 times. The comparison uses different configurations, 32 GPUs for Vera Rubin and 48 GPUs for GB300, and both meet the response-speed condition of at least 50 tokens per second per user.
| Item | Vera Rubin | GB300 |
|---|---|---|
| GPUs | 32 | 48 |
| Concurrent queries | 1,024 | 48 |
| Total throughput (tokens/sec) | 41,414.2 | 2,210.4 |
| Operating power | 100.89 kW | 60 kW |
| Average per-user generation speed (tokens/sec) | 50.4 | 56.2 |
| Time to first token (P90) | 4,673.9 ms | 2,305.1 ms |
| Verification status | Provisional, awaiting MLCommons verification | MLCommons verified |
Measurements from the Standardized division, Available category of MLPerf Endpoints v1.0, as stated in NVIDIA's announcement. The common condition is a per-user generation speed of at least 50 tokens per second.
Throughput per unit of power is calculated by dividing the system's total tokens per second by its operating power in kW. Vera Rubin consumes more power than GB300, but it raises the system's overall output by greatly increasing the number of queries processed simultaneously. The result compares how much processing the systems can perform per unit of power consumed, after setting a minimum per-user generation speed.
There is, however, a difference in how long users wait for the first response. Time to first token (TTFT) at P90 indicates that 90% of queries received their first token within that time. In this measurement, Vera Rubin recorded 4,673.9 ms and GB300 recorded 2,305.1 ms, so Vera Rubin's wait is longer. Even with better throughput per unit of power, that alone does not allow an assessment of the wait before a response begins, deployment costs, or the total energy needed for the same workload.
In addition, the Vera Rubin figures being compared are provisional results that have not completed MLCommons verification, so they are at a different stage of publication from GB300's verified results. The 10x-plus ratio should be treated as a comparison of throughput per unit of power under these specific conditions.
Software and real operating conditions also matter in deployment decisions
Even on the same hardware, performance changes with software updates. After its official submission results were finalized, NVIDIA also published re-measured results applying newer software to the same hardware configurations. For example, when training Llama 3.1 8B on 72 Vera Rubin GPUs, the official submission recorded 7.162 minutes, while the re-measurement recorded 6.140 minutes. However, this re-measured value predates MLCommons' verification and review and should be distinguished from the official results.
When considering an AI platform, it is important first to compare training time for the target model at the number of GPUs you can actually secure. If you plan to add GPUs, the system configuration, including networking, and the scaling efficiency as the count grows are also factors in the decision. For inference services, what matters is how much throughput can be secured with the available power while meeting the expected number of concurrent queries and response-speed requirements.
The results show that Vera Rubin recorded shorter training times than GB300 on certain training tasks and obtained provisional figures indicating high power efficiency in inference. Going forward, in addition to the verified MLPerf Endpoints results, measurements that align the actual GPU scale, model, software, networking configuration, and response conditions will be what informs judgments about performance and cost at deployment.
