NVIDIA claims up to 2x training performance per GPU in Rubin’s MLPerf debut
The November 16 results pair faster training with a power-efficiency claim for AI agents. That inference comparison uses different system configurations and still awaits verification.
NVIDIA’s MLPerf Training v6.1 results suggest Vera Rubin can improve training throughput per GPU while retaining efficiency as workloads scale: its DeepSeek-V3 run cut training time from 16 minutes to about four when GPU count rose from 256 to 1,024. Separately, NVIDIA claims over 10 times the output per watt on an agentic-workload test, but that result uses different system configurations and is still awaiting MLCommons verification. For infrastructure buyers, the results offer performance signals, not yet a fully verified, like-for-like basis for comparing systems.
01
NVIDIA reported roughly 97% scaling efficiency for Rubin’s DeepSeek-V3 run when increasing the allocation from 256 to 1,024 GPUs.
02
The largest DeepSeek-V3 submissions used 10,240 GPUs: NVIDIA ran GB300 systems, while Microsoft Azure used GB200 systems.
03
On Kimi K3, Rubin sustained 41,414.2 tokens per second at 100.89 kW; GB300 sustained 2,210.4 at 60 kW.
NVIDIA says its Vera Rubin NVL72 systems delivered up to twice the training performance per GPU of GB300 NVL72 in their MLPerf Training debut. Published November 16, 2026, the announcement also pitches a larger gain in AI-agent output per watt—but that inference result remains provisional and awaits MLCommons verification.
Training results put scaling in focus
The new MLPerf Training v6.1 results measure how quickly systems train models to a target accuracy. NVIDIA said its platform led every benchmark and was the only one represented across the full suite, covering large language models, image generation, recommendation systems and reinforcement-learning post-training.
The up-to-2x figure compares performance per GPU with NVIDIA’s GB300 NVL72 system. A separate DeepSeek-V3 training result shows how Rubin handled a larger allocation of processors: moving from 256 to 1,024 GPUs reduced training time from 16 minutes to about four, with roughly 97% scaling efficiency, according to NVIDIA.
That is close to the ideal outcome of quadrupling processors and cutting elapsed time to one-quarter. NVIDIA attributes its coordination across GPUs to NVLink connections within systems, InfiniBand and Ethernet networking between systems, and optimized software. Adding processors only shortens training if they can continue working together efficiently.
The largest training submissions in the round used other NVIDIA systems. NVIDIA ran DeepSeek-V3 on 10,240 GPUs using GB300 hardware; Microsoft Azure matched that GPU count using GB200 systems. Those entries provide scale context, rather than extending Rubin’s reported 256-to-1,024-GPU result to a larger machine.
The agent-efficiency claim has a different baseline
The same announcement included NVIDIA’s first MLPerf Endpoints submission, citing Kimi K3 results published November 9. Endpoints evaluates agentic coding and enterprise workflows and checks that agents complete tasks correctly. Unlike training time, this comparison concerns the output systems can sustain while serving users within a power budget.
NVIDIA reported more than 10 times the throughput per watt for Rubin versus GB300 on Kimi K3. The comparison uses a shared minimum response-generation rate of 50 tokens per second per user. Tokens are the output units counted by the test; the efficiency measure divides total token throughput by operating power.
The disclosed test configurations
Vera Rubin: 32 accelerators, 1,024 concurrent queries and 41,414.2 tokens per second, drawing 100.89 kilowatts.
GB300: 48 accelerators, 48 concurrent queries and 2,210.4 tokens per second, drawing 60 kilowatts.
Those are system-level configurations with different accelerator counts and numbers of simultaneous queries—not a per-GPU comparison like the training headline. Both exceeded the response-generation threshold: Rubin averaged 50.4 tokens per second per user, versus 56.2 for GB300. The claim therefore concerns aggregate output per watt at a minimum interactive pace, not responses arriving 10 times faster.
Nor were all responsiveness measures identical. The reported 90th-percentile time to the first token was 4,673.9 milliseconds for Rubin and 2,305.1 milliseconds for GB300. The common comparison point was sustained output per user, not an identical wait before output began.
The next formal step for the Rubin Endpoints result is MLCommons verification and peer review. NVIDIA says the GB300 result is already verified; Rubin’s figures remain subject to change through that process.
Sources
blogs.nvidia.comNVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance
Reader comments
Newest comments first. Replies stay oldest first.