Supermicro Extends Its NVIDIA Blackwell System Portfolio

Supermicro, Inc. reports expansion of its NVIDIA Blackwell system portfolio with the new 4U DLC-2 liquid-cooled NVIDIA HGX B200 system and the debut of an air-cooled 8U front I/O system. Tailored to large-scale AI training and cloud-scale inference workloads, Supermicro's systems streamline maintenance of air- or liquid-cooled AI infrastructure and are made to support coming NVIDIA HGX B300 platforms, Supermicro says.

"Supermicro's DLC-2 enabled NVIDIA HGX B200 system leads our portfolio to achieve greater power savings and faster time to online for AI Factory deployments," says Charles Liang, CEO and president, Supermicro. "Supermicro's extensive portfolio now can offer precisely optimized NVIDIA Blackwell solutions to a diverse range of AI infrastructure environments, whether deploying into an air- or liquid-cooled facility."

Supermicro's DLC-2, a comprehensive cooling architecture, features the following: 

• Up to 40% data center power savings
• Faster time to deployment and reduced time to online by providing an data center scale liquid cooling solution
• Up to 40% reduced water consumption with warm water cooling at an inlet temperature of up to 45°C
• Up to 98% system heat capture by liquid-cooling CPUs, GPUs, DIMMs, PCIe switches, VRMs, power supplies 
• Enabling quiet data center operation at a noise level as low as 50dB

Supermicro offers a broad portfolio of NVIDIA HGX B200 solutions with the two new front I/O systems and six rear I/O systems. 

"Advanced infrastructure is accelerating the AI industrial revolution for every industry," says Kaustubh Sanghani, vice president of GPU product management at NVIDIA. "Based on the latest NVIDIA Blackwell architecture, Supermicro's new front I/O B200 systems equip enterprises to deploy and scale AI at unprecedented speed—delivering breakthrough innovation, efficiency, and operational excellence."

The system's 8 high-performance 400G NVIDIA ConnectX-7 NICs and 2 NVIDIA Bluefield-3 DPUs are moved to the front of the system to allow for configuration of networking cables, storage drive bays, and management all from the cold aisle. 

In addition to system architecture improvements, Supermicro has fine-tuned components for AI data center workloads. Upgraded memory expansion with 32 DIMM slots delivers for system memory configuration. Large system memory complements the NVIDIA HGX B200's HBM3e GPU memory.

All Supermicro 4U liquid-cooled systems and 8U or 10U air-cooled systems are optimized for NVIDIA HGX B200 8-GPU with each GPU connected via 5th-generation NVLink at 1.8TB/s. NVIDIA's Blackwell platform delivers faster real-time inference performance and faster training for LLMs compared to the Hopper generation of GPUs. The new front I/O systems with dual-socket CPUs supporting up to 350W Intel Xeon 6 6700 Series processors deliver performance for AI workloads, Supermicro reports.

The 4U front I/O liquid-cooled system features front-accessible NICs, DPUs, storage, and management components. It utilizes dual-socket Intel Xeon 6700 Series processors with P-cores up to 350W and NVIDIA HGX B200 8-GPU configuration (180GB HBM3e per GPU).

The system supports 32 DIMMs with up to 8TB capacity at 5200MT/s or up to 4TB at 6400MT/s DDR5 RDIMM, plus 8 hot-swap E1.S NVMe storage drive bays and 2 M.2 NVMe boot drives. Network connectivity includes 8 single-port NVIDIA ConnectX-7 NICs or NVIDIA BlueField-3 SuperNICs and two dual-port NVIDIA BlueField-3 DPUs.

The 8U front I/O air-cooled system shares the same front-accessible architecture and core specifications while providing a solution for AI factories without liquid-cooling infrastructure. It features a compact 8U form factor (compared to Supermicro's 10U system) with a reduced-height CPU tray while maintaining the full 6U-height GPU tray.

For more information, visit https://www.supermicro.com/en/accelerators/nvidia 

Sources: Press materials received from the company and additional information gleaned from the company’s website.