NVIDIA DGX H200 AI System

The NVIDIA DGX H200 AI System (MPN: 936-24287-0000-001) is NVIDIA's flagship purpose-built AI infrastructure appliance, integrating eight NVIDIA H200 SXM Tensor Core GPUs interconnected via NVLink 4.0 and NVSwitch 3.0 to form a unified 1,128 GB pool of HBM3e memory. Each H200 GPU delivers 989 TFLOPS of FP8 Tensor Core performance and 3.35 TB/s of memory bandwidth per GPU, enabling the system to sustain the extreme data throughput required by the largest generative AI models, large language models, and deep learning training workloads at scale.

Built on the Hopper GPU architecture, the DGX H200 provides a fully optimized, pre-integrated software and hardware stack through the NVIDIA DGX software suite, including NVIDIA Base Command Manager for cluster orchestration and access to the NVIDIA NGC catalog of containerized AI frameworks. The system connects to external storage and networking via eight NVIDIA ConnectX-7 400Gb/s InfiniBand/Ethernet adapters, enabling high-bandwidth, low-latency fabric integration into multi-node DGX SuperPOD or third-party AI cluster environments.

The DGX H200 is engineered for the most demanding enterprise AI workloads, including training and fine-tuning of foundation models with hundreds of billions of parameters, high-throughput generative AI inference serving, and compute-intensive scientific simulation. Its rack-optimized 10U form factor, integrated dual-rail high-bandwidth networking, and validated thermal and power infrastructure make it a production-ready appliance for enterprise data centers deploying AI at scale.

Ideal for

• Training and fine-tuning large language models (LLMs) and multimodal foundation models with hundreds of billions of parameters
• High-throughput generative AI inference serving for enterprise-scale deployments of models such as GPT-class and diffusion-based architectures
• Accelerated drug discovery and molecular dynamics simulation requiring large in-memory datasets and high FP64/FP32 compute throughput
• Enterprise recommender system training on massive sparse embedding tables that benefit from the expanded 141 GB per-GPU HBM3e memory capacity
• Multi-node distributed deep learning research and development integrated into DGX SuperPOD or InfiniBand-connected AI fabric clusters
• Retrieval-augmented generation (RAG) pipeline development and deployment combining large embedding models with real-time inference at scale

Technical specifications

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