2026-11-04 · Fonte:Unite AI

First MLPerf Inference preview results show NVIDIA's next-generation rack claiming up to 3.7 times the throughput of GB300 NVL72 on the Qwen3-VL benchmark.
NVIDIA published its MLPerf Inference v6.1 submission results on September 16, 2026, and the headline entry was the first MLPerf Inference preview submission for the company's Vera Rubin NVL72 system. According to figures released alongside the results, the platform posted up to 3.7 times higher throughput than NVIDIA's GB300 NVL72 system on the Qwen3-VL benchmark, marking the first standardized performance data point for the next-generation rack.
The numbers were produced under MLPerf Inference: Datacenter, a benchmark suite maintained by the MLCommons Association that measures how fast systems process inputs and produce results using a trained model. Entries are published under MLCommons' definitions, and the preview designation flags this as the Vera Rubin NVL72's first appearance in the suite rather than a standard submission.
For organizations planning large-scale AI deployments, the generational framing is the notable part. The 3.7x figure is a direct rack-to-rack comparison against the GB300 NVL72, and it was recorded on the Qwen3-VL workload in particular. The 'up to' qualifier signals a peak result on a single benchmark rather than an average across the full suite, so buyers should treat it as an early indicator of the new system's throughput profile rather than a guaranteed uplift for every model and serving scenario.
MLPerf Inference: Datacenter exists precisely to enable comparisons of this kind: identical workloads, disclosed configurations and a common yardstick for how quickly systems turn trained models into completed outputs. NVIDIA's September 16, 2026 release therefore sets an early reference point for the Vera Rubin NVL72, and the up-to-3.7x margin over GB300 NVL72 on Qwen3-VL will be the figure to verify — and to beat — as subsequent MLPerf rounds bring fuller submissions.
Tags: NVIDIA · Vera Rubin NVL72 · GB300 NVL72 · MLPerf Inference · MLCommons · Benchmarks
萬安算交所 AIXX · AIXX · Exchange verificado · Ir ao mercado