Supermicro NVIDIA GB200 NVL72 48U Rack Solution

The Supermicro NVIDIA GB200 NVL72 48U Rack Solution (SRS-GB200-NVL72) is a rack-scale AI supercomputing system built on NVIDIA's Blackwell architecture, integrating 36 Grace CPU modules and 72 Blackwell B200 GPU dies within a single 48U rack enclosure. Each GB200 superchip pairs two B200 GPUs with one Grace Arm-based CPU via NVLink-C2C interconnect, delivering a unified, high-bandwidth memory and compute fabric. The system leverages NVLink Switch technology to unite all 72 GPUs into a single high-speed NVLink domain at 1.8 TB/s per GPU of NVLink bandwidth, enabling the entire rack to operate as one logical GPU for large-scale model training and inference.

The GB200 NVL72 rack is engineered for the most demanding frontier AI workloads, including trillion-parameter large language model (LLM) training, mixture-of-experts (MoE) inference at scale, and scientific simulation. The Grace CPU subsystem contributes 480 GB of LPDDR5X memory per rack accessible at CPU-class bandwidth, while each B200 GPU delivers 192 GB of HBM3e memory, resulting in approximately 13.8 TB of total HBM3e capacity across the rack. Supermicro's MGX mechanical design framework enables the dense integration of compute, networking, and power delivery into the 48U form factor with liquid cooling as the primary thermal solution.

The system is designed for direct liquid cooling (DLC) to support its exceptional thermal load, making it suited for modern data centers equipped with liquid cooling infrastructure. High-speed networking is supported through NVIDIA Quantum-X800 InfiniBand or Spectrum-X800 Ethernet fabric connectivity, enabling multi-rack cluster scaling for hyperscale AI training clusters. The SRS-GB200-NVL72 is positioned for enterprise data centers, national AI labs, cloud service providers, and large-scale HPC facilities across UAE, GCC, EMEA, and APAC regions seeking to deploy next-generation AI infrastructure.

Ideal for

• Frontier large language model (LLM) pre-training and fine-tuning at trillion-parameter scale, leveraging the unified 72-GPU NVLink domain as a single logical compute unit
• High-throughput generative AI inference serving for production enterprise deployments requiring maximum tokens-per-second throughput with reduced latency
• Scientific high-performance computing (HPC) simulations including climate modeling, molecular dynamics, and computational fluid dynamics benefiting from the combined Grace CPU and Blackwell GPU architecture
• Mixture-of-experts (MoE) model training and inference where the high-capacity HBM3e memory pool across 72 GPUs accommodates extremely large sparse model states
• Multi-modal AI workload execution combining vision, language, and structured data processing within a single rack-scale unified memory and compute environment
• Sovereign AI and national-scale AI infrastructure deployments in GCC, EMEA, and APAC data centers requiring certified, rack-integrated GPU systems with enterprise-grade support

Technical specifications

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