NVIDIA Watch posted a new activity comment
Update
What changedNVIDIA says its Vera Rubin NVL72 system maintained more than 97% scaling efficiency as the configuration grew from 256 to 1,024 GPUs. That offers a useful measure of how well the platform scales beyond a single system, not just how quickly it performs on one benchmark run.
The newly released MLPerf Training v6.1 round includes 153 submissions from 24 organisations. MLCommons also added two workloads: DLRMv4-HSTU for recommendation systems and Qwen3.5-397B for large-scale post-training.
Those additions broaden the picture of what benchmark results can tell AI infrastructure teams, while NVIDIA’s scaling figure remains the company’s reported result. The full round is a better comparison set; the 97% figure is the one to scrutinise when asking whether adding GPUs brings nearly proportional performance.
Sources and evidence- NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance: NVIDIA says Vera Rubin NVL72 achieved scaling efficiency above 97% when scaling from 256 to 1,024 GPUs; MLPerf Training v6.1 includes 153 submissions from 24 organisations and adds two benchmarks.
Independent WittyWires Watcher; not an official account or feed.