The NVIDIA DGX H100 System is an enterprise AI supercomputer designed to deliver end-to-end performance for Generative AI, LLM training/inference, and high-performance computing (HPC). Built with 8× NVIDIA H100 Tensor Core GPUs and NVIDIA’s high-speed NVLink/NVSwitch fabric, DGX H100 provides a complete, optimized platform for running the most demanding AI workloads—from data preparation and training to large-scale inference and multi-node scaling.
Key Highlights
- 8× NVIDIA H100 GPUs (640GB total GPU memory) for large models and high-throughput training/inference.
- 4× NVIDIA NVSwitch delivering 7.2 TB/s bidirectional GPU-to-GPU bandwidth for fast multi-GPU scaling inside the system.
- NVLink per GPU: 18 connections / 900 GB/s bidirectional GPU-to-GPU bandwidth for ultra-low latency, high-bandwidth GPU communication.
- 10× NVIDIA ConnectX-7 400Gb/s NICs providing up to 1 TB/s peak bidirectional network bandwidth for cluster scale-out.
- Dual Intel Xeon 8480C (56 cores each) for strong host compute and I/O throughput.
- 2TB system memory (DDR5) for large datasets and demanding pipelines.
- Enterprise NVMe storage: 2× 1.92TB NVMe M.2 (RAID1) for OS + 8× 3.84TB NVMe U.2 for high-speed local data cache (common config).
