en.Wedoany.com Reported - Super Micro Computer, Inc. has introduced Data Center Building Block Solutions (DCBBS) blueprints based on the NVIDIA Vera Rubin NVL72 and NVIDIA HGX Rubin NVL8 platforms. These blueprints are designed for gigawatt-scale AI data center deployments, starting from a single scalable unit containing 1152 GPUs that can expand to virtually any scale. Supermicro's DCBBS blueprints provide an end-to-end holistic solution from design to delivery, backed by a dedicated expert team covering the entire deployment lifecycle. The solution integrates compute, storage, networking, advanced liquid cooling, power distribution, and on-site infrastructure to accelerate the time-to-live for large-scale liquid-cooled AI factories.
Supermicro's DCBBS blueprints address the practical implementation challenges of the world's most advanced AI infrastructure. The NVIDIA Vera Rubin platform significantly increases the performance density of AI factories, doubling speeds across multiple computing domains. NVIDIA's latest reference architecture precisely defines the ideal content of a 1152-GPU scalable unit, while Supermicro's DCBBS blueprints define the specific steps for successful deployment. Supermicro has a proven track record of deploying the world's largest liquid-cooled AI factories with over 100,000 GPUs.
Customers planning to build or retrofit AI factories commonly face a rigid constraint: available power. The DCBBS blueprints for the NVIDIA Vera Rubin NVL72 provide a balanced bill of materials within specified power ranges (from 5MW to 1GW), with appropriately proportioned cooling capacity, power delivery, compute nodes, management nodes, high-performance storage nodes, context memory storage platform nodes, and networking to eliminate performance bottlenecks caused by network oversubscription, power capacity limitations, or thermal throttling.
These blueprints cover Supermicro's complete end-to-end process for executing large-scale AI projects at record speed. On-site facility surveys are conducted by Supermicro's dedicated team, analyzing the physical site's compatibility with deployment requirements, including loading docks, data hall dimensions, clearance, floor plans, and floor load ratings. Project design and planning incorporate all critical details into construction plans tailored to customer requirements and facility constraints. Supermicro defines the correct combination of DCBBS components, such as cooling solutions: for fully direct liquid cooling-compatible facilities, rack-level cooling distribution units (CDUs) up to 1.8MW; for sites without facility water infrastructure, liquid-to-air sidecar units; while also developing in-rack CDU options based on 52U rack configurations, and supplemental rear-door heat exchanger options for high-ambient-temperature environments. During the solution integration phase with comprehensive on-site services, most work is completed at its U.S. manufacturing facilities, including racking, stacking, and cabling within each rack. Supermicro validates functionality through testing processes exceeding industry standards, covering system-level (L10) and cluster-level (L11) multi-node testing. Dedicated teams manage logistics for on-site components such as CDUs, cooling towers, and power infrastructure, including coordination with third-party vendors selected by the customer. Integrated delivery services and on-site integration encompass rack placement, power and cooling connections, network cabling, system commissioning, software stack installation, and on-site validation. In terms of support, services, and software, a range of ongoing on-site options are available, including fastest 4-hour on-site response for mission-critical uptime requirements, and support for integration with infrastructure management tools such as Supermicro SuperCloud Composer and SuperCloud Director, enabling unified control from bare-metal management to multi-tenant workload orchestration, while incorporating software stacks like NVIDIA AI Enterprise and NVIDIA Run:ai.
The NVIDIA Vera Rubin platform has the potential to deliver transformative generational performance improvements, but requires a repeatable and reliable approach for successful deployment. Supermicro ensures alignment with the latest NVIDIA reference architectures, giving customers confidence that their deployments are consistent with the NVIDIA cloud partner ecosystem. The scalable unit at the core of the Supermicro DCBBS blueprints provides 1152 NVIDIA Rubin GPUs with 331TB of HBM4 GPU memory. Compared to NVIDIA Blackwell, the Vera Rubin generation doubles GPU memory bandwidth, GPU-to-GPU NVLink bandwidth, and per-GPU network bandwidth, providing the architectural foundation for training and inferencing top-tier AI models with trillions of parameters. Its supporting infrastructure includes: an advanced direct liquid cooling stack (DLC-2), encompassing a 5MW cooling tower, four rack-level CDUs up to 1.8MW, 16 vertically mounted cooling distribution manifolds, 576 direct-to-chip copper cold plates, and Supermicro SMC PG25-A coolant with excellent chemical and thermal stability, along with 200kW and 500kW liquid-to-air options; power distribution infrastructure, from medium-voltage transformers to low-voltage distribution, rack-level power shelves, and battery backup units (BBUs), with each Vera Rubin NVL72 rack equipped with four 110kW power shelves and redundant 18.3kW power supply units, supporting battery energy storage systems (BESS) for instant switchover backup power; optimized 48U and 52U racks; 16 compute racks optimized for the NVIDIA Vera Rubin NVL72 and NVIDIA HGX Rubin NVL8 platforms; 6 network racks (4 for compute, 2 for convergence), supporting NVIDIA Spectrum-X Ethernet or NVIDIA Quantum-X800 InfiniBand, with compute fabric speeds up to 1.6TB/s, and silicon photonics network options based on co-packaged optics (CPO) eliminating pluggable transceivers; 4 high-performance storage racks based on the Supermicro Petascale server platform for NVMe-tier application storage and model training checkpoints; plus 2 context memory storage platform racks, optimized for long-context inference, agent working memory, and retrieval workloads. For more information, visit supermicro.com/vera-rubin.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
