Supermicro has launched a brand new Information Middle Constructing Block Options (DCBBS) Blueprint for HPC based mostly on the NVIDIA Vera Rubin NVL4 platform. The announcement extends the corporate’s DCBBS structure technique past enterprise AI deployments to scientific computing environments, offering a reference framework for deploying large-scale HPC and AI infrastructure.
The brand new blueprint follows the NVIDIA Vera Rubin NVL72 and NVIDIA HGX Rubin NVL8 DCBBS designs unveiled earlier this 12 months and applies the identical built-in strategy to analysis establishments, nationwide laboratories, and supercomputing facilities. Supermicro’s DCBBS framework combines compute, networking, liquid cooling, energy distribution, and facility infrastructure right into a pre-engineered deployment mannequin that reduces implementation complexity and accelerates cluster deployment.
Supermicro President and CEO Charles Liang mentioned AI is changing into a elementary element of contemporary scientific analysis. He famous that organizations able to quickly deploying superior infrastructure will likely be positioned to drive future discoveries. He highlighted the corporate’s expertise constructing large-scale liquid-cooled GPU clusters as a key issue behind the brand new blueprint.
Designed for Converged HPC and AI Workloads
The blueprint is constructed across the rising convergence of conventional HPC simulation and AI-accelerated computing. Analysis organizations more and more mix double-precision simulation workloads with machine studying and AI fashions to shorten growth cycles and speed up scientific discovery throughout fields equivalent to local weather modeling, drug discovery, supplies science, and power analysis.
Supermicro positions the NVIDIA Vera Rubin NVL4 platform as a basis for these hybrid workloads, combining large-scale GPU acceleration with CPU assets and high-performance networking. The corporate notes that its deployment methodology relies on expertise gained from constructing liquid-cooled supercomputing clusters containing greater than 100,000 GPUs.
Finish-to-Finish Deployment Methodology
The DCBBS Blueprint outlines the deployment course of Supermicro makes use of for large-scale liquid-cooled infrastructure tasks.
Tasks start with on-site facility assessments that consider loading dock entry, knowledge corridor dimensions, flooring loading capabilities, and current energy and cooling infrastructure. The ensuing knowledge is used to develop a site-specific deployment plan.
System integration is carried out in Supermicro manufacturing amenities earlier than cargo, together with rack meeting, cabling, and each system-level (L10) and cluster-level (L11) validation testing. As soon as delivered, deployment providers embody rack set up, community integration, energy and cooling connections, commissioning, and operational validation.
Supermicro additionally provides ongoing help providers, together with on-site response choices meant for mission-critical analysis and manufacturing environments.
Vera Rubin NVL4 Scalable Unit Structure
The blueprint is organized round a repeatable NVIDIA Vera Rubin NVL4 Scalable Unit that may be replicated to create clusters starting from roughly 3.2MW deployments to amenities approaching gigawatt scale.
Every scalable unit consists of eight liquid-cooled compute racks utilizing customized 52U enclosures. The configuration helps 288 NVIDIA Vera Rubin NVL4 nodes, incorporating as much as 1,152 NVIDIA Rubin GPUs and 576 NVIDIA Vera CPUs. Particular person racks function inside a 362kW energy envelope.
Cooling is offered by way of Supermicro’s DLC-2 direct liquid cooling structure. Every scalable unit contains three in-row cooling distribution items, every rated at as much as 1.8MW, in a 2+1 redundant configuration. The cooling design incorporates direct-to-chip chilly plates, vertical coolant distribution manifolds, and Supermicro’s PG25-A coolant formulation.
Networking relies on NVIDIA Quantum-X800 InfiniBand infrastructure, offering the low-latency, high-bandwidth interconnect required for distributed HPC and AI workloads. Liquid-cooled networking configurations are additionally out there.
Energy supply is dealt with by eight 72kW energy cabinets per compute rack, whereas twin top-of-rack administration switches present out-of-band monitoring and management.
Rapid Availability for GB200 NVL4 Deployments
Along with the Vera Rubin-based structure, Supermicro mentioned configurations based mostly on NVIDIA GB200 NVL4 can be found instantly for organizations in search of near-term deployment choices.
The announcement displays the continued business shift towards built-in rack-scale infrastructure designs that mix compute, networking, energy, cooling, and administration into pre-validated deployment fashions. As AI and HPC workloads proceed to converge, distributors are more and more specializing in decreasing deployment complexity whereas supporting increased rack densities and liquid-cooled environments.
