NVIDIA Vera Rubin NVL72 debuts on MLPerf Inference — up to 3.7x GB300

NVIDIA’s company blog said Vera Rubin NVL72’s first MLPerf Inference v6.1 preview submission delivered up to 3.7x better throughput than GB300 NVL72 on Qwen3-VL, and up to 2.5x on DeepSeek-R1, with GB300 also showing 99% multi-rack scaling efficiency.

HARDWARE desk — same-day company primary that puts Vera Rubin NVL72 on the public MLPerf Inference scoreboard against GB300, giving buyers a concrete throughput and scaling comparison for AI inference racks.

Company claim: Vera Rubin NVL72 delivered up to 3.7x better throughput than GB300 NVL72 on Qwen3-VL across offline, server, and interactive scenarios. Throughput here means how many tokens or queries the rack can serve. Qwen3-VL is a vision-language model in the suite — it reads text and images. Offline, server, and interactive are the three load patterns MLPerf uses: a batch of work, a live stream of requests, and a snappier interactive stream. NVIDIA says those Qwen3-VL runs used vLLM with NVIDIA Dynamo, an open inference stack. File 3.7x and those three scenarios as NVIDIA’s, via MLPerf entries 6.1-0106 and 6.1-0074.

Company claim: on DeepSeek-R1, Vera Rubin NVL72 delivered up to 2.5x higher throughput than GB300 NVL72. DeepSeek-R1 is a reasoning model in the same suite. NVIDIA says those runs used TensorRT-LLM, the company’s library for serving large language models. File 2.5x as NVIDIA’s, via the same named entries.

Company claim: a 288-GPU, four-rack GB300 NVL72 DeepSeek-R1 submission reached 99% scaling efficiency in the offline scenario versus a single 72-GPU rack. NVL72 is NVIDIA’s 72-GPU rack-scale system — seventy-two of the company’s AI chips wired as one cabinet. Scaling efficiency means how much extra throughput you get when you add more GPUs. 99% means four racks did almost four times the work of one. File 288 GPUs, four racks, and 99% as NVIDIA’s, via MLPerf entries 6.1-0073 and 6.1-0074.

Company claim: software optimizations lifted GB300 NVL72 Qwen3-VL performance up to 1.6x versus the prior v6.0 results. File 1.6x as NVIDIA’s software-stack claim on the same blog. Extra kernel-fusion and cache-precision talk stays in Sources.

Company notes on the same post: Nebius also submitted Vera Rubin NVL72 preview results. NVIDIA submitted Jetson AGX Thor on a new Edge-Agentic benchmark with Qwen3.6-27B — a smaller model meant for devices at the edge of the network, not a full data-center rack. The company says 19 partners participated broadly. File Nebius, Jetson AGX Thor, and the 19-partner count as NVIDIA’s. Partner names stay in Sources.

Plain English for the rest of the card: MLPerf = industry benchmark suite run by MLCommons. NVL72 = NVIDIA’s 72-GPU rack-scale system. Vera Rubin = NVIDIA’s next-generation platform after Blackwell / GB300. GB300 = the current NVIDIA rack Vera Rubin is being scored against. throughput = how many tokens or queries the rack can serve. inference = running a trained model to answer a query, not teaching a new one. preview submission = on the scoreboard, not yet in the available-to-buy MLPerf bucket. scaling efficiency = how much extra throughput you get when you add more GPUs. Qwen3-VL = the suite’s vision-language model. DeepSeek-R1 = the suite’s reasoning model.