NVIDIA Vera Rubin NVL72 Debuts in MLPerf Inference v6.1 with 3.7x Throughput Leap
In its MLPerf Inference v6.1 debut, NVIDIA Vera Rubin NVL72 rack-scale AI supercomputer demonstrated up to 3.7x higher throughput over GB300 NVL72 on Qwen3-VL and DeepSeek-R1.
NVIDIA published official preview results for its next-generation Vera Rubin NVL72 rack-scale computing platform in the industry-standard MLPerf Inference v6.1 benchmarks. Demonstrating massive generational performance acceleration, the Vera Rubin system delivered up to 3.7x greater throughput than the current flagship GB300 NVL72 on the Qwen3-VL multimodal vision-language model and up to 2.5x higher throughput on the DeepSeek-R1 complex reasoning benchmark across server and offline deployment scenarios.
3.7x Throughput Gain on Frontier Reasoning and Multimodal Workloads
The MLPerf Inference v6.1 round marks the public benchmarking debut of the Vera Rubin platform on state-of-the-art foundation models. By coupling extreme silicon compute density with architectural optimizations for generative token throughput, Vera Rubin NVL72 established wide margins over current-generation hardware. On Qwen3-VL, the system sustained up to 3.7x more concurrent inferencing queries without latency degradation, while DeepSeek-R1 reasoning workloads saw up to a 2.5x throughput multiplier, showcasing the platform efficacy in handling complex chained thought verification and high-parameter reasoning.
Benchmark Highlights & System Performance
- Multimodal Vision Throughput: Up to 3.7x higher inference throughput compared to NVIDIA GB300 NVL72 on Qwen3-VL.
- Reasoning Acceleration: Up to 2.5x throughput gain on DeepSeek-R1 reasoning benchmarks across offline and server testing modes.
- Linear Cluster Scaling: A 288-GPU deployment spanning four GB300 NVL72 racks achieved 99% linear scaling efficiency.
- Full-Stack Software Boost: Software and compiler optimizations within TensorRT-LLM contributed an additional 1.6x speedup over MLPerf v6.0 baselines.
Vera CPU, NVLink 6, and HBM4 Architectural Innovations
The Vera Rubin NVL72 architecture unites 72 Rubin GPUs with NVIDIA proprietary Vera CPUs—purpose-built to orchestrate autonomous agentic pipelines and data movement. The system integrates 20.7 TB of ultra-high-speed HBM4 memory delivering record-setting bandwidth across a liquid-cooled single-rack footprint. Inter-GPU communication is driven by NVLink 6 switch fabrics providing 3.6 TB/s bidirectional bandwidth per GPU and 260 TB/s aggregate rack backplane bandwidth, transforming the entire 72-GPU assembly into a single unified accelerator domain.
Product Profile Sheet
| Feature | Details |
|---|---|
| Architecture | Vera CPU + Rubin GPU + NVLink 6 Switches |
| Interconnect Bandwidth | 3.6 TB/s per GPU (260 TB/s Aggregate NVLink 6) |
| MLPerf Benchmark Suite | MLPerf Inference v6.1 (September 2026) |
| Memory Pool | 20.7 TB HBM4 Memory (Rack-Scale Unified) |
| Peak Throughput Uplift | Up to 3.7x vs. GB300 NVL72 (Qwen3-VL) |
| Primary Source | NVIDIA Technical Blog / MLCommons |
| System Model | NVIDIA DGX Vera Rubin NVL72 |