NVIDIA GB200 NVL36 Rack-Scale System 6912GB HBM3e

The NVIDIA GB200 NVL36 is a rack-scale AI computing system integrating 36 Blackwell GPUs and 18 Grace CPUs within a single NVLink-connected rack unit, delivering unified high-bandwidth memory and compute resources across the entire chassis. Built on the NVIDIA Blackwell GPU architecture, each GB200 Superchip pairs two B200 GPUs with one Grace CPU via NVLink-C2C interconnect, achieving up to 900 GB/s chip-to-chip bandwidth. The rack system aggregates 6,912 GB of HBM3e memory across all GPUs, providing an unprecedented pooled memory capacity for training and inferencing extremely large AI models that previously required multi-rack deployments.

The GB200 NVL36 delivers up to 1.4 exaflops of AI compute (FP4) across the full rack, with NVLink Switch System fabric enabling all 36 GPUs to communicate at 1.8 TB/s of all-to-all GPU bandwidth. This tight integration eliminates traditional PCIe bottlenecks and allows the rack to operate as a single logical unit for distributed AI workloads. The system is purpose-built for the most demanding large language model (LLM) training runs, mixture-of-experts (MoE) architectures, and real-time inference serving at scale, where inter-GPU communication latency and memory capacity are the primary constraints.

Designed for modern AI-first data centers, the GB200 NVL36 uses a liquid-cooled architecture to manage the substantial thermal output of a full rack of Blackwell Superchips, enabling high-density deployment within standard data center footprints. The system integrates with NVIDIA's full software stack including CUDA, cuDNN, TensorRT, and NEMO frameworks, and supports NVIDIA Confidential Computing for enterprise security requirements. Organizations across cloud service providers, enterprise AI research, financial services, and life sciences rely on this platform to consolidate AI infrastructure and reduce the total number of nodes required for frontier model workloads.

Ideal for

• Training frontier-scale large language models (LLMs) with hundreds of billions to trillions of parameters, leveraging the 6,912 GB pooled HBM3e memory to keep model weights and optimizer states on-chip across the full rack
• High-throughput LLM inference serving for enterprise AI applications, using the NVLink-connected GPU pool to minimize token generation latency at scale for internal or customer-facing deployments
• Mixture-of-experts (MoE) model training and inference where large aggregate memory and ultra-low-latency all-to-all communication between GPUs are critical to maintaining expert routing efficiency
• Scientific simulation and HPC workloads in pharmaceutical drug discovery, climate modeling, and genomics that require tightly coupled GPU compute with large shared memory spaces
• Enterprise AI model fine-tuning pipelines for domain-specific foundation models in financial services, healthcare, and manufacturing, where rapid iteration over large proprietary datasets is required
• Sovereign AI and national AI infrastructure deployments requiring dense, rack-scale compute capacity with support for confidential computing and enterprise-grade data governance

Technical specifications

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