The Pulse
NVIDIA reports Vera Rubin gains in MLPerf Training debut
NVIDIA says Vera Rubin NVL72 delivered up to twice the training performance per GPU of its GB300 system in the latest MLPerf round. A separate agentic inference comparison showed more than 10 times the throughput per watt, but the Vera Rubin result remains provisional.

NVIDIA’s Vera Rubin NVL72 delivered up to twice the training performance per GPU of its GB300 NVL72 system in its first MLPerf Training round, the company said. The company also reported more than 10 times the throughput per watt for Vera Rubin in a separate agentic inference comparison, though that result is provisional.
The MLPerf Training results were published November 16; the MLPerf Endpoints results NVIDIA cites, which test systems on agentic coding and enterprise workflows, were published November 9. They measure different kinds of work: training a model to a target accuracy and running an AI system to produce responses.
Training at larger scale
The latest MLPerf Training suite spans large language models, image generation, recommender systems and reinforcement learning post-training. The company said its platform was the only one with results across every benchmark in the round.
For the DeepSeek-V3 training benchmark, NVIDIA reported that Vera Rubin NVL72 scaled from 256 to 1,024 GPUs at about 97% efficiency. Increasing the GPU count fourfold reduced training time from 16 minutes to about four minutes.
NVIDIA also ran the benchmark on 10,240 GPUs using GB300 systems, which it said was the largest scale in the round. Microsoft Azure submitted a run at the same scale using 10,240 NVIDIA GPUs in GB200 systems.
MLPerf Training measures the time needed to train a model on a specified dataset to a quality target. MLCommons says it repeats runs, discards the highest and lowest results and averages the remainder; its published guidance puts result variation at very roughly plus or minus 2.5% for imaging benchmarks and very roughly plus or minus 5% for other benchmarks.
A provisional inference comparison
The Kimi K3 comparison used throughput per kilowatt at a shared minimum end-to-end average interactivity of 50 tokens per second per user. The provisional Vera Rubin run used 32 accelerators, reached 41,414.2 tokens per second in system throughput and had operating power of 100.89 kW.
The GB300 run used 48 accelerators, reached 2,210.4 tokens per second and had operating power of 60 kW. NVIDIA said the GB300 result had been verified by MLCommons, while the Vera Rubin result was awaiting verification and could change after peer review.
MLCommons defines system throughput as the total tokens produced per second across users, while interactivity measures the rate at which a response streams back to each user. Endpoints results also track concurrent queries and time to first token, and can be normalized for power consumption.
Other submissions
NVIDIA said 20 organizations submitted results using its platform in the training round, including ASUS, Azure, Cisco, CoreWeave, Dell Technologies, Google Cloud, HPE, Oracle and the University of Florida.
Vera Rubin had appeared in a different MLPerf suite before this training round. In September, NVIDIA published preview submissions for MLPerf Inference covering the DeepSeek-R1 and Qwen3-VL benchmarks.
Sources
Companies in this story
Related stories
- Google reports Ironwood TPU results in MLPerf training tests
- MLPerf Training adds agentic and recommender tests in v6.1
- NVIDIA Partner Firebird Secures License for 45,000 GPUs in Kazakhstan
- NVIDIA adds AI workload inventory to GPU platform runtime
- NVIDIA’s 120B Nemotron 3.5 Super VL is available on SageMaker