CONFIRMEDDate confidence: CONFIRMEDStated as firmHow firmly the source words the date: confirmed, scheduled, estimated, delayed or unverified.Show all CONFIRMED pins86% CONFIDENCEOverall confidence: 86%How well the pin's source and references back up its dates.Weighted average of how firmly 6 references, the source included, support the pin's start and end times; a reference counts half as much for every 180 days older than the newestShow all pins at 75% confidence or better
MLCommons[4] published the MLPerf Inference v6.1 results on 16 September 2026 - its GlobeNewswire release is datelined 'SAN FRANCISCO, Sept. 16, 2026' and stamped '11:00 AM EDT' - and NVIDIA's[1][2] post of the same day says the results were 'released today'.- Start
- Sep 16, 202695% CONFIDENCEmlcommons.org
NVIDIA Vera Rubin NVL72 Debuts in MLPerf Inference v6.1
In its first peer-reviewed benchmark, NVIDIA's Vera Rubin NVL72 rack posted up to 3.7x the throughput of GB300 NVL72 on Qwen3-VL and 2.5x on DeepSeek-R1, submitted in the preview category of MLCommons' MLPerf Inference v6.1 round.
- MLCommons published the MLPerf Inference v6.1 results on 16 September 2026, the first peer-reviewed numbers for NVIDIA's Vera Rubin NVL72, which was entered in the preview category[3][4]
- On Qwen3-VL, Vera Rubin NVL72 delivered up to 3.7x the throughput of GB300 NVL72 across offline, server and interactive scenarios using vLLM with NVIDIA Dynamo, and up to 2.5x on DeepSeek-R1 using TensorRT-LLM[1]
- Nebius, the only other Vera Rubin submitter, ran DeepSeek R1 on 36 GPUs across nine nodes and posted 16,427 tokens/s per GPU in the server scenario[6]
- A 288-GPU GB300 NVL72 submission across four racks reached 99% scaling efficiency on DeepSeek-R1, and software work lifted GB300 NVL72 on Qwen3-VL by up to 1.6x over v6.0[1]
- The round drew a record 30 submitting organizations and added End-to-End RAG and Edge Agentic Inference tests; the best DeepSeek R1 per-accelerator server result was 5.7x a year earlier[4]
- NVIDIA's claim of 30x over GB300 NVL72 on SemiAnalysis AgentX comes from preview testing outside MLPerf and is not verified by MLCommons[1][3]
- Results were published at 11:00 a.m. EDT through MLCommons' GlobeNewswire release[5]
Notable features
- One rack holds 72 Rubin GPUs and 36 Vera CPUs, with ConnectX-9 SuperNICs and BlueField-4 DPUs[2]
- 20.7 TB of HBM4 at 1,400 TB/s, 3,168 custom Olympus CPU cores and up to 54 TB of LPDDR5X per rack[2]
- Sixth-generation NVLink and NVLink Switch, which NVIDIA says gives 10x higher packet rates and 3x lower latency than off-the-shelf Ethernet[1]
- NVFP4 precision and enhanced Tensor Cores and Transformer Engine speed both prefill and decode[1]
- NVIDIA rates it at one-tenth the cost per million tokens of GB200 NVL72[2]