NVIDIA GB200 NVL72 Rack System via Dell PowerEdge

The NVIDIA GB200 NVL72 Rack System, integrated and delivered through Dell PowerEdge infrastructure, represents the most advanced AI supercomputing platform available in a single rack form factor. Built on NVIDIA's Blackwell architecture, the NVL72 configuration unifies 36 Grace CPU modules and 72 Blackwell B200 GPU dies interconnected via fifth-generation NVLink, forming a single coherent 72-GPU compute fabric with a combined 1.4 exaflops of AI performance at FP4 precision. The system integrates NVIDIA's NVLink Switch technology to enable full all-to-all GPU communication at 130TB/s of aggregate NVLink bandwidth across the entire rack, effectively treating 72 GPUs as one massive unified compute engine.

Each GB200 Superchip pairs one Grace CPU with two Blackwell GPU dies, connected via NVLink-C2C at 900GB/s bidirectional bandwidth, eliminating traditional PCIe bottlenecks and enabling CPU-GPU memory coherency. The rack system incorporates 13.5TB of unified HBM3e memory across all 72 GPU dies, with each GPU die providing 192GB of HBM3e at 8TB/s memory bandwidth. The Grace CPUs contribute additional LPDDR5X capacity, and the coherent memory architecture allows both CPU and GPU threads to access the full memory pool without explicit data migration. Liquid cooling is employed throughout the rack via a direct liquid cooling (DLC) manifold system, as the thermal envelope of the NVL72 exceeds the capacity of air-cooled solutions.

Delivered through Dell's enterprise integration and PowerEdge rack ecosystem, the GB200 NVL72 is designed for organizations deploying frontier-scale AI model training, large language model inference at scale, and high-performance scientific simulation. The system ships as a complete, rack-integrated solution including NVLink Switch Trays, compute trays, power distribution, and liquid cooling infrastructure, enabling rapid deployment into data center environments equipped with rear-door heat exchangers or facility-level liquid cooling. Omnixon Global makes this platform accessible to enterprise and hyperscale buyers across the UAE, GCC, EMEA, and APAC regions.

Ideal for

• Frontier-scale large language model (LLM) training, enabling organizations to train models with hundreds of billions to trillions of parameters within a single rack's unified memory and compute fabric
• High-throughput LLM inference serving for real-time generative AI applications, leveraging the 13.5TB unified HBM3e pool to hold extremely large models entirely in GPU memory without model sharding across nodes
• Climate and physical sciences simulation requiring tightly coupled CPU-GPU workloads that benefit from NVLink-C2C memory coherency and Grace CPU LPDDR5X capacity alongside GPU compute
• Drug discovery and molecular dynamics simulation at scale, utilizing FP8 and FP4 Tensor Core throughput for deep learning-driven protein structure prediction and generative molecular design
• Computer vision and multimodal AI model training for autonomous systems and industrial inspection, taking advantage of the NVL72's transformer engine and second-generation FP8 support
• Enterprise AI infrastructure consolidation, replacing multi-rack GPU clusters with a single NVL72 rack to reduce data center footprint, networking complexity, and inter-node communication overhead

Technical specifications

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