NVIDIA says its new training system delivered up to 2x performance per GPU over existing products. Model teams gain a new option to cut training waits with the same GPU count. The company says it is supplying systems through cloud infrastructure partners.

1. Generational differences with the same model and GPU count

The company says Vera Rubin NVL72 recorded the fastest training times in 6 of 7 benchmarks in its MLPerf Training v6.1 debut.

According to MLCommons, the test measures the time to train a model to a target quality using specified data. The Closed division compares hardware and software using the same model as the reference implementation.

The biggest generational gap was on DeepSeek-V3 671B, which requires extensive communication. NVIDIA reports 4.14 minutes with 1,024 GPUs in 16 racks, frames housing servers. Below is its three-generation comparison at the same model and GPU count.

Table: Training times across three generations at the same scale

Conditions: MLPerf Training v6.1 Closed division, DeepSeek-V3 671B, 1,024 GPUs in each NVL72 system

Unit: minutes

Training time by generation with the same model and GPU countGB200 NVL72: Training time 12.36; GB300 NVL72: Training time 8.03; Vera Rubin NVL72: Training time 4.14GB200 NVL7212.36GB300 NVL728.03Vera Rubin NVL724.14
Training time by generation with the same model and GPU count
Generational differences with the same model and GPU count
SystemTraining time
GB200 NVL7212.36
GB300 NVL728.03
Vera Rubin NVL724.14

NVIDIA reports about 2x performance per GPU on this model versus optimized GB300 at the same GPU count. Other models also improved, but by 1.4–1.6x.

2. Peak speed and scaling efficiency are different records

NVIDIA says GB300 trained the same model in 1.38 minutes using 10,240 GPUs in 160 racks, the round's fastest time. This distinguishes improving per-GPU efficiency from finishing sooner by using more GPUs.

The company says Vera Rubin exceeded 97% scaling efficiency on this model when increasing from 256 to 1,024 GPUs and from 4 to 16 racks. This means it retained most of the speedup proportional to added resources.

Within a rack, NVLink 6 connects 72 GPUs and 36 CPUs, providing bandwidth of 3.6TB/s per GPU and 260TB/s per rack. Bandwidth is the amount of data transferred per second. Together with inter-rack networking, it supports large-scale computing.

3. A new post-training test and the role of CPUs

The newly added Qwen3.5 397B GRPO is a reinforcement-learning post-training test in which a model tries actions and learns from evaluations. NVIDIA's article reports 10.44 minutes for Vera Rubin and 13.78 minutes for GB300 on 1,024 GPUs.

Submissions also used two inference software stacks. At the same scale, Vera Rubin recorded 10.44 minutes with TensorRT-LLM and 10.94 minutes with vLLM.

Post-training needs both GPU computing and CPU resources for simulations that test actions in modeled real-world situations. NVIDIA says 256 CPUs in one Vera CPU rack support over 22,500 reinforcement-learning environments simultaneously.

4. Software improvements also benefit existing systems

After the submission deadline, NVIDIA reran each model with a new software build, keeping hardware and settings unchanged. All times below are company reference results. In each paragraph, Vera Rubin comes first and GB300 second.

Llama 3.1 8B took 6.140 and 9.211 minutes on 72 GPUs. gpt-oss 20B took 8.080 and 13.151 minutes on 64 GPUs.

DLRMv4 took 1.628 and 2.940 minutes on 128 GPUs. GB300's official submission time was 3.871 minutes, showing that the new build also improved the previous generation.

The company's reinforcement-learning performance multiples use 64 GPUs, while its article's training times use 1,024. Reruns lack MLCommons validation and MLPerf review, and may differ from official results in measurement or implementation.

5. Shorter training times versus cluster costs

Shorter training allows more experiments in the same period. That benefit requires operating large GPU clusters and inter-rack links, plus CPU simulation resources for post-training. Teams can size deployments using target-model training times and full system configurations.

Vera Rubin's debut suggests a practical selection criterion: the total system scale needed to cut waits for the target model, rather than the highest performance multiple.