Supermicro has announced the expansion of its NVIDIA Blackwell architecture portfolio with the introduction and shipment of new 4U and 2-OU (Open Compute Project, or OCP) liquid-cooled NVIDIA HGX B300 systems. These systems are designed for hyperscale and AI factory data center deployments, enabling high GPU density and improved energy efficiency through advanced direct liquid-cooling technology.
The 4U liquid-cooled NVIDIA HGX B300 system is built for standard 19-inch EIA racks and enables up to 64 GPUs per rack. The system utilizes Supermicro’s Direct Liquid-Cooling 2 (DLC-2) technology, capturing up to 98 percent of system heat through liquid-cooling. This enables large-scale AI factory deployments to achieve power efficiency and straightforward system servicing for dense training and inference clusters.
The 2-OU OCP liquid-cooled NVIDIA HGX B300 system targets OCP Open Rack V3 (21-inch) infrastructure. This configuration supports up to 144 GPUs per rack, enabled by a compact eight-GPU system architecture. The rack-scale design incorporates blind-mate manifold connections, modular GPU/CPU tray architecture, and modern liquid-cooling solutions. Each node can drive eight NVIDIA Blackwell Ultra GPUs at up to 1,100 W thermal design power. A fully populated ORV3 rack hosts up to 18 nodes with 144 GPUs, and Supermicro highlights scalability through the integration of NVIDIA Quantum-X800 InfiniBand switches and in-row coolant distribution units supporting up to 1.8 MW per rack. In cluster configurations, a SuperCluster unit can consist of up to 1,152 GPUs using eight compute racks, three networking racks, and two in-row coolant distribution units.
Both systems provide 2.1 TB of HBM3e GPU memory per system and double compute fabric network throughput up to 800 Gb/s using NVIDIA ConnectX-8 SuperNICs with NVIDIA Quantum-X800 InfiniBand or Spectrum-4 Ethernet. These features are designed to accelerate industrial-scale AI workloads such as agentic AI systems, foundation model training, and multimodal inference in data center environments.
Supermicro reports its DLC-2 technology stack can deliver up to 40 percent power savings and reduce water consumption by supporting 45°C warm water operation, eliminating the need for chilled water and compressors in data centers. The company delivers these new rack-scale solutions as validated, tested L11 and L12 systems to speed up deployment for hyperscale, enterprise, and government customers.
The new HGX B300 systems join Supermicro’s existing range of NVIDIA Blackwell platforms, including the NVIDIA GB300 NVL72, HGX B200, and RTX PRO 6000 Blackwell Server Edition. The systems are validated for compatibility with NVIDIA networking and software, allowing customers to scale AI infrastructure from a single node to full data center-scale factories.
Source: Supermicro
