tekno.BALI ISLAND
EN
All articles
Computer5 min read

SC26 Chicago: NVIDIA Vera Rubin NVL72 Breaks MLPerf v6.1 Records and ASUS AI Factory Expansion

cover-nvidia-vera-rubin-nvl72-sc26

This week, the global Computer and high-performance computing (HPC) infrastructure industry gathered in Chicago for Supercomputing 2026 (SC26). If previous years were dominated by discussions of individual chip shortages, the main spotlight at this year's SC26 shifted entirely to rack-scale architecture, which treats an entire server cabinet as a single, interconnected giant computer unit.

Two major announcements stole the main stage: the competitive debut of the NVIDIA Vera Rubin NVL72, which immediately set new records in industry MLPerf testing, and ASUS accelerating the mass shipment of the AI POD XA VR721-E3 liquid-cooled rack system while expanding local AI access to consumer PCs and workstations.


1. Architectural Breakdown of the NVIDIA Vera Rubin NVL72

Unlike traditional servers that limit inter-GPU communication within a single chassis, the Vera Rubin NVL72 platform is designed as a unified rack-scale computing domain. According to the official DGX Vera Rubin NVL72 datasheet, a single rack integrates 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs (based on Arm-compatible custom cores) working coherently.

The training needs of trillion-parameter models and AI agent reasoning demand ultra-fast data delivery to compute cores. To address this, the Vera Rubin architecture brings three key foundations:

  • Massive HBM4 Capacity: Each Rubin GPU is equipped with up to 288 GB of HBM4 memory, generating a total of 20.7 TB of HBM4 memory in a single rack with an aggregate bandwidth reaching the petabyte-per-second scale (approximately 1.4–1.6 PB/s depending on configuration).
  • Sixth-Generation NVLink: Powered by nine NVLink 6 switch trays, the scale-up communication path between GPUs in a single rack reaches 216 TB/s to 260 TB/s, eliminating bottlenecks as thousands of model parameters exchange gradients.
  • High-Density Scale-Out Networking: To connect multiple racks into a supercomputer cluster, the system combines 144 ConnectX-9 SuperNICs and 18 BlueField-4 DPUs to support high-speed InfiniBand and Spectrum-X Ethernet traffic.

2. Clean Sweep in MLPerf Training v6.1 and Endpoints v1.0

The new architecture's claims were immediately put to the test in the official November 2026 MLPerf benchmark round, where the Vera Rubin NVL72 made its premiere against previous-generation systems.

a. Training Performance Doubled per GPU

In the official MLPerf Training v6.1 results, NVIDIA noted that the Vera Rubin NVL72 is capable of delivering up to 2x the performance per GPU compared to the thoroughly optimized GB300 NVL72 (Blackwell Ultra) system. In addition to sweeping all testing categories, the system demonstrated a scaling efficiency above 97% when clusters were expanded from 256 to 1,024 GPUs.

b. AI Agent Inference Energy Efficiency

Meanwhile, in the MLPerf Endpoints v1.0 agent inference tests, the focus shifted to the throughput-to-power consumption ratio. Under modern deep reasoning and multimodal model workloads (such as Qwen3-VL and DeepSeek variants), the Vera Rubin NVL72 posted an efficiency gain of over 10x throughput per kilowatt compared to previous-generation rack systems—a crucial indicator given that electricity and cooling costs are now the largest constraints on AI data center expansion.


3. ASUS Accelerates AI POD XA VR721-E3 Shipments & Local AI Expansion

On the ecosystem partner manufacturing side, ASUS utilized the SC26 stage to demonstrate mass production readiness. Through the ASUS Vera Rubin NVL72 shipment announcement, the company confirmed accelerated production and full shipments in the fourth quarter of 2026 for its flagship unit, the ASUS AI POD XA VR721-E3.

a. Full Liquid Cooling and Six-Layer Validation

Considering that a single XA VR721-E3 rack supports four 110 kW power shelves with peak power requirements of up to 227 kW, conventional air cooling systems are no longer sufficient. Therefore, ASUS implemented a 100% liquid-cooled architecture and relies on automated L10/L11 assembly lines equipped with optical inspection and precise tightening. Each rack must also pass a six-layer validation framework—covering thermal, firmware, and cross-GPU network integrity tests—before being shipped to cloud providers and enterprises.

b. Bringing AI to DIY PCs, Workstations, and Desktops

Interestingly, ASUS's strategy does not stop at giant server rooms. Beyond the data center segment, ASUS also announced the expansion of the Local AI ecosystem to PCs, spanning do-it-yourself (DIY) PC components, gaming devices, creator workstations, and professional desktops. This move is aimed at reducing the Total Cost of Ownership (TCO) for organizations and creators who want to run AI agent workflows locally or in a hybrid manner without having to rely entirely on cloud computing rentals.


4. 2027–2028 HPC Roadmap: Scientific Computing & Quantum

SC26 also affirmed that the Vera and Rubin architectures are not built solely for commercial large language models. In the NVIDIA HPC roadmap announcement series, two upcoming lines for the scientific research community were revealed:

  • Vera 192-CPU HPC Rack (Early 2027): A dense configuration highlighting custom Arm CPU computing capacity for physics simulation workloads, climate modeling, and large-memory workloads.
  • Rubin HPC GPU (2028): A specialized high-precision scientific computing variant designed to bring over 50,000 CUDA cores to advance next-generation scientific research and quantum simulations.

Conclusion

The SC26 event in Chicago demonstrated how the boundary between single chips and data center infrastructure has dissolved. Through the Vera Rubin NVL72, NVIDIA proved that a 2x per-GPU performance surge and 10x per-kilowatt power efficiency can only be achieved through a rack-scale design that integrates HBM4 memory, NVLink 6, and liquid cooling. At the same time, the accelerated production of the ASUS AI POD XA VR721-E3 and the push to bring Local AI acceleration to workstations and DIY PCs signal that this new generation of computing power is ready to flow from giant data centers straight to the desks of creators and developers worldwide.

Wilan

A regular contributor to Bali Island Tekno who actively shares knowledge about technology, programming, and the world of software engineering.

Read next