Supermicro SYS-421GE-NBRT-LCC 4U Liquid-Cooled HGX B200 Server

The Supermicro SYS-421GE-NBRT-LCC is a 4U rack-mounted server built around the NVIDIA HGX B200 8-GPU baseboard, delivering the full compute capability of NVIDIA Blackwell architecture with an advanced direct liquid cooling (DLC) thermal subsystem. Unlike traditional air-cooled HGX deployments, the LCC (Liquid-Cooled Configuration) variant routes coolant directly through cold plates on the GPUs, CPUs, and high-TDP VRMs, dramatically reducing facility cooling overhead and enabling sustained operation at peak TDP levels that would otherwise throttle in air-cooled racks. The system is designed to slot into modern hot-aisle/cold-aisle liquid-cooled row infrastructure or direct-to-chip manifold systems, making it a natural fit for next-generation AI data centers targeting PUE values well below 1.2.

At its computational core, the server houses eight NVIDIA B200 SXM GPUs interconnected via fifth-generation NVLink with an aggregate 1.44 PB/s GPU-to-GPU bandwidth across the NVLink fabric. Two 5th-generation Intel Xeon Scalable processors (Granite Rapids) provide host-side compute, supported by 32 DDR5 DIMM slots capable of accommodating up to 8 TB of system memory. PCIe 5.0 connectivity across multiple expansion slots ensures that NVMe storage, high-speed networking NICs, and offload accelerators operate without bandwidth starvation. The HGX B200 baseboard itself integrates 1.44 TB of HBM3e memory across the eight GPUs, providing the vast on-accelerator memory headroom required for running frontier-scale large language models and multimodal foundation models at full precision or with reduced quantization overhead.

Targeted at AI training, large-scale inference, scientific simulation, and high-performance data analytics, the SYS-421GE-NBRT-LCC represents Supermicro's most thermally efficient HGX B200 platform. The liquid cooling loop supports facility inlet water temperatures up to 45°C in specific configurations, allowing operators to leverage free cooling or economizer cycles for a greater portion of the year. Management is handled through Supermicro's IPMI 2.0-compliant BMC with Redfish API support, enabling integration into OpenBMC-based fleet orchestration, Kubernetes-based GPU cluster managers, and enterprise monitoring stacks. The chassis supports tool-less GPU module serviceability designed to minimize downtime during component replacement in production environments.

Ideal for

• Large language model (LLM) pre-training and fine-tuning at frontier scale, leveraging 1.44 TB HBM3e and NVLink fabric for multi-node tensor and pipeline parallelism
• High-density inference serving for generative AI applications where sustained GPU TDP and low-latency response require liquid-cooled sustained clock frequencies
• Multi-modal foundation model training combining vision, language, and audio modalities that demand both high memory capacity and extreme inter-GPU bandwidth
• Scientific computing and HPC simulation workloads in climate modeling, computational fluid dynamics, and molecular dynamics that benefit from the Blackwell architecture's FP8 and FP16 tensor core throughput
• Liquid-cooled AI cluster buildouts in co-location facilities with direct-to-chip cooling infrastructure seeking to maximize GPU density per cabinet without breaching power-per-rack limits
• Enterprise AI platform consolidation where operators need to replace multiple older GPU server generations with fewer, thermally stable nodes running containerized multi-tenant inference workloads

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.