Supermicro SuperServer SYS-422GA-NBRT-LCC (NVIDIA HGX B200 8-GPU)

The Supermicro SuperServer SYS-422GA-NBRT-LCC is a 4U rack-mount platform built around the NVIDIA HGX B200 8-GPU baseboard, pairing eight NVIDIA Blackwell B200 SXM6 GPUs with dual 5th-generation Intel Xeon Scalable or AMD EPYC host processors. Each B200 GPU delivers up to 20 petaFLOPS of FP4 AI compute and 192 GB of HBM3e memory per device, yielding a staggering 1.536 TB of aggregate GPU memory across the node—sufficient to host frontier-scale large language models without model parallelism across nodes. The platform leverages NVLink 5.0 at 1.8 TB/s all-to-all GPU bandwidth, ensuring near-linear scaling for transformer-based training workloads that demand ultra-low inter-GPU latency.

The 'LCC' (Liquid-Cooled Configuration) suffix designates direct liquid cooling on both the GPU baseboard and the host processors, enabling sustained Thermal Design Power operation at full GPU boost clocks in high-density data center deployments where air cooling cannot satisfy the thermal envelope of Blackwell-class accelerators. The system supports PCIe 5.0 x16 host-to-GPU connectivity alongside NVIDIA NVSwitch fabric internal to the HGX baseboard, providing both high-bandwidth GPU-direct storage paths and coherent GPU collective operations. Storage is provisioned through front-accessible NVMe U.2 bays and optional OCP 3.0 mezzanine NICs, enabling GPUDirect RDMA over 400GbE InfiniBand NDR or RoCEv2 Ethernet fabrics.

Targeted at hyperscalers, national laboratories, cloud service providers, and enterprise AI centers of excellence, the SYS-422GA-NBRT-LCC is optimized for GPT-class LLM pre-training, multimodal foundation model fine-tuning, high-throughput inference serving, and scientific simulation workloads including molecular dynamics and climate modeling. Supermicro's Building Block Solutions architecture allows customers to integrate the system into standard 19-inch rack infrastructure with centralized CDU (Coolant Distribution Unit) manifolds, facilitating rack-scale liquid cooling deployments at 40–60 kW per rack without supplemental in-row air cooling.

Ideal for

• Large language model pre-training (GPT-4 class and beyond) requiring hundreds of billions of parameters resident in pooled HBM3e memory across a single node
• High-throughput generative AI inference serving with continuous batching frameworks such as NVIDIA TensorRT-LLM, maximizing tokens-per-second at sub-100ms SLA
• Multimodal foundation model fine-tuning combining vision encoders and language decoders where NVLink 5.0 all-reduce bandwidth eliminates gradient synchronization bottlenecks
• Computational fluid dynamics and climate simulation leveraging FP64 Blackwell tensor cores for mixed-precision scientific workloads at national-lab scale
• Reinforcement Learning from Human Feedback (RLHF) pipelines where reward model and policy model co-reside on shared GPU memory to eliminate inter-node communication overhead
• AI-driven drug discovery and protein structure prediction running ensemble simulations that benefit from 1.536 TB aggregate GPU memory for batching thousands of molecular configurations simultaneously

Technical specifications

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