NVIDIA Vera Rubin NVL72 Debuts in MLPerf Inference v6.1 with 3.7x Qwen3-VL Throughput

NVIDIA's Vera Rubin NVL72 made its preview debut in MLPerf Inference v6.1, posting up to 3.7x higher Qwen3-VL throughput than the GB300 NVL72 across all tested scenarios. The company attributes the gains to enhanced Tensor Cores and a Transformer Engine tuned for both prefill and decode phases. For DeepSeek-R1, the Vera Rubin NVL72 delivered up to 2.5x higher throughput than the GB300 NVL72, using the TensorRT-LLM library. These results come from NVIDIA's own submissions, with partner Nebius also contributing preview results that showed excellent performance, according to NVIDIA.

Software and Scaling Gains

The MLPerf Inference v6.1 results also highlight software improvements that benefit the existing GB300 NVL72. NVIDIA reports that optimizations in v6.1 gave the GB300 NVL72 up to 1.6x performance over v6.0 on Qwen3-VL. Post-v6.1 software updates added further gains on GPT-OSS-120B and DLRMv3, though these are unverified. The GB300 NVL72 submission used 288 GPUs across four racks and reached 99% scaling efficiency, with throughput growing nearly linearly from a single-rack baseline. NVIDIA points to continuous software improvements, including lower KV cache precision, kernel fusion, better kernels, and disaggregated serving with vLLM and Dynamo, as drivers of these gains.

Architecture and Precision

NVIDIA highlights several architectural features behind the Vera Rubin NVL72's performance. NVFP4 precision reduces memory usage for weights, attention, and KV cache, boosting throughput with minimal quality loss. Disaggregated serving separates prefill and decode, and expert parallelism is used for MoE layers. The NVL72 scale-up relies on sixth-generation NVLink and NVLink Switch, which NVIDIA claims offers 10x packet rate and 3x lower latency than Ethernet. On the SemiAnalysis AgentX benchmark, the Vera Rubin NVL72 was 30x faster than the GB300 NVL72 in preview results.

Edge and Ecosystem

Beyond the large-scale systems, NVIDIA's Jetson AGX Thor ran TensorRT Edge-LLM on the Edge-Agentic benchmark with Qwen3.6-27B. The company also notes that the Vera Rubin NVL72 rack delivers more tokens, serves more users, and lowers cost per token versus a GB300 rack. The GB300 NVL72 achieved 9x higher throughput and 7.5x lower latency than a single node on a WAN 2.2 text-to-video benchmark, hitting 0.65 720p fps at 5.7 seconds per video. NVIDIA's partner ecosystem includes 19 partners, eight of which use multi-node Blackwell NVL72 systems. MLPerf Inference v6.1 closed-division results were retrieved on Sep 16, 2026, with Vera Rubin entries 6.1-0106 and 6.1-0074, and GB300 entries 6.1-0073 and 6.1-0074. NVIDIA describes its platform as purpose-built for performance, scaling efficiency, and software velocity in AI inference.