NVIDIA's Vera Rubin debuts on MLPerf Training with the fastest time on six of seven benchmarks, about 2x per GPU on DeepSeek-V3
Synopsis
NVIDIA's Vera Rubin architecture entered MLPerf Training v6.1 for the first time, and its NVL72 platform posted the fastest training time on six of the seven benchmarks, up to 2x the highly optimized GB300 NVL72 baseline. On the 671-billion-parameter DeepSeek-V3 workload, Vera Rubin NVL72 cut training from 8.03 to 4.14 minutes at 1,024 GPUs, while the round's fastest absolute result, 1.38 minutes, came from a 10,240-GPU GB300 NVL72. NVIDIA reports above 97% overall scaling efficiency for Vera Rubin from 256 to 1,024 GPUs.
Interpretation
NVIDIA's Vera Rubin architecture entered the MLPerf Training benchmark for the first time, and its NVL72 rack-scale platform delivered the fastest training time on six of the seven benchmarks, up to 2x the company's own highly optimized GB300 NVL72 baseline. The architecture had no prior MLPerf training results; the comparison is against an already tuned GB300 NVL72 configuration on the same models at the same GPU counts. On the 671-billion-parameter DeepSeek-V3 benchmark, Vera Rubin NVL72 trained in 4.14 minutes using 1,024 GPUs across 16 racks, while GB300 NVL72 ran the identical 1,024-GPU configuration in 8.03 minutes, close to 2x per GPU and the largest generational gain of the round.
Across the remaining benchmarks the lead falls between 1.4x and 1.6x, spanning dense models, mixture-of-experts models, a diffusion model and a recommender, so the gain is not confined to one model shape. In the same round, GB300 NVL72 took the fastest absolute result on the seventh benchmark, DeepSeek-V3, at 1.38 minutes across 10,240 GPUs, so the two platforms play different roles. Benchmark by benchmark, the ratios are 1.4x for Qwen3.5 397B GRPO at 64 GPUs, 1.6x for gpt-oss 20B at 64 GPUs, 1.5x for Llama 2 70B LoRA at 72 GPUs, 1.6x for Llama 3.1 8B at 72 GPUs, 1.5x for FLUX.1 at 72 GPUs and 1.4x for DLRMv4 at 128 GPUs.
Those times rest on co-designed networking inside and outside the rack: 72 GPUs and 36 CPUs connect through a copper NVLink 6 spine within the rack, and ConnectX-9 SuperNICs carry traffic out to InfiniBand or Ethernet. Every iteration of a mixture-of-experts model produces a burst of all-to-all communication, so fabric efficiency shows up directly in time to train. The NVLink 6 domain delivers 3.6 TB/s per GPU and 260 TB/s per rack, and both external fabrics use adaptive routing and congestion control to keep all-to-all traffic behind the compute layer; on DeepSeek-V3, scaling from 256 to 1,024 GPUs (4 to 16 racks) held above 97% efficiency.
Perspective
These numbers are NVIDIA's submissions to MLPerf Training v6.1 as listed by MLCommons, and the comparison is against NVIDIA's own GB300 NVL72 baseline rather than other vendors' systems; the separately listed reruns on a newer software build are unverified and have not been through MLPerf review.
NVIDIA says Vera Rubin's CUDA software stack is already highly mature at debut and expects the initial performance advantages to keep unlocking as the stack matures; that expectation has not yet been tested by later MLPerf rounds. The reruns on a newer software build are explicitly labeled unverified and may use measurement methodologies or workload implementations inconsistent with the MLPerf specification.
