Supermicro GB200 NVL72 Rack-Scale System (SRS-GB200NVL72)

The Supermicro SRS-GB200NVL72 is a purpose-built rack-scale AI infrastructure system based on NVIDIA's Grace Blackwell Superchip architecture, housing 72 NVIDIA Blackwell GPUs and 36 NVIDIA Grace ARM-based CPUs within a single NVLink-connected fabric. The system is engineered as a unified compute domain rather than a collection of discrete servers, enabling all 72 GPUs to operate as one cohesive accelerated compute engine with fifth-generation NVLink providing 1.8 TB/s bidirectional bandwidth per GPU. This architecture eliminates traditional PCIe bottlenecks and allows the full rack to function as a single logical unit for the largest generative AI training and inference workloads.

The system leverages NVIDIA's NVLink Switch technology to create a fully non-blocking, all-to-all GPU interconnect across all 72 GPUs in the rack. Each GB200 NVL72 module pairs two Blackwell B200 GPUs with one Grace CPU via chip-to-chip NVLink, delivering 900 GB/s CPU-to-GPU bandwidth. The combined rack delivers a peak FP8 compute density exceeding 11.5 exaFLOPS, with 13.5 TB of aggregate HBM3e memory providing the capacity required for training and serving frontier-scale large language models with hundreds of billions to trillions of parameters.

Supermicro's implementation incorporates a direct liquid cooling architecture that handles the 120 kW+ rack-level thermal load with precision manifold distribution, ensuring sustained GPU boost performance across all 72 accelerators simultaneously. The system ships as a pre-integrated, factory-validated rack unit designed to minimize on-site integration complexity for hyperscale and enterprise AI data centers, supporting direct integration with high-performance storage fabrics, InfiniBand or Ethernet networking, and Supermicro's full lifecycle management tooling via IPMI and Redfish-compliant BMC interfaces.

Ideal for

• Training frontier large language models with 100B+ parameters where GPU-to-GPU bandwidth and memory capacity are the primary scaling constraints
• High-throughput generative AI inference serving for production LLM deployments requiring maximum tokens-per-second density per rack
• Multimodal foundation model development combining vision, language, and speech modalities at scale within a single unified compute domain
• AI supercomputing cluster construction where multiple GB200 NVL72 racks are interconnected via InfiniBand NDR400 for petascale to exascale training jobs
• Scientific simulation and digital twin workloads in climate modeling, drug discovery, and physics simulation that require tightly coupled GPU collaboration
• Enterprise AI factory deployments consolidating multiple model training pipelines on a shared, high-utilization rack-scale infrastructure

Technical specifications

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