NVIDIA DGX H100 ( 8×H100 SXM5 AI Supercomputing Platform )

Description

The NVIDIA DGX H100 is a powerful AI workstation built for heavy AI workloads, deep learning training, and large-scale model inference. It comes with 8 NVIDIA H100 GPUs, each with 80GB of super-fast HBM3 memory, connected through high-speed NVLink, making it perfect for running large AI models quickly.

It uses 2 Intel Xeon Platinum 8480+ CPUs (56 cores each) and 2TB of DDR5 memory, providing plenty of processing power and memory for demanding tasks. Storage is handled by 8 NVMe SSDs (3.84TB each), giving a total of ~30TB fast storage for AI datasets.

The DGX H100 supports popular AI frameworks such as vLLM, Megatron, DeepSpeed, HuggingFace Transformers, TensorRT-LLM, and Triton Inference Server, making it flexible for research and production.

Networking is extremely fast with high-speed ConnectX-7 NICs and NVLink + Infiniband for connecting multiple systems together. The system also includes secure boot, TPM 2.0, RAID support, and redundant 3000W power supplies for reliability. It runs NVIDIA DGX OS 6, ready for container-based AI workflows.

Key Features:

Product Description

The NVIDIA DGX H100 represents the gold standard for enterprise AI infrastructure, delivering an unprecedented leap in performance, scalability, and security. Purpose-built as the engine for large language models (LLMs), recommender systems, and complex high-performance computing (HPC) workloads, the DGX H100 is a fully integrated hardware and software solution designed to remove the barriers to AI innovation. It empowers data scientists, researchers, and developers to tackle the world’s most challenging computational problems, moving from concept to deployment faster than ever before.

At its core, the DGX H100 is powered by the revolutionary NVIDIA Hopper™ architecture. This system combines eight NVIDIA H100 Tensor Core GPUs with the high-speed connectivity of 4th generation NVLink, creating a single, massive GPU with 640GB of total HBM3 memory. This unified memory space is critical for training enormous models that would otherwise be impossible to fit on a single GPU. The platform is turnkey, arriving with the NVIDIA AI Enterprise software suite, providing a secure, stable, and supported environment for building production-ready AI. From drug discovery and climate science to generative AI and autonomous vehicle development, the DGX H100 is the essential tool for organizations leading the next wave of artificial intelligence.

Technical Specifications

Frequently Asked Questions (FAQ)

Q1: What is the primary difference between the DGX H100 and the previous DGX A100? The DGX H100 is a significant generational leap. Its key advantages include the new NVIDIA Hopper GPU architecture, faster HBM3 memory (vs. HBM2e on A100), 4th generation NVLink with double the bandwidth, PCIe Gen5 support, and the new Transformer Engine, which dramatically accelerates AI models like LLMs. This results in performance gains of up to 6x for AI training and 30x for AI inference compared to the DGX A100.

Q2: Who is the ideal user for a DGX H100 system? The DGX H100 is designed for large enterprises, leading research institutions, and AI-focused startups that need to train and deploy state-of-the-art AI models at scale. It is ideal for teams working on natural language processing, drug discovery, autonomous systems, financial modeling, and other computationally intensive AI applications.

Q3: What software is included with the DGX H100? The system comes with the NVIDIA DGX OS, which is built on the NVIDIA AI Enterprise software suite. This is an end-to-end, cloud-native suite of AI and data analytics software, optimized and supported by NVIDIA. It includes access to popular frameworks like PyTorch and TensorFlow, as well as NVIDIA-specific tools like TensorRT-LLM, Triton Inference Server, and NeMo Megatron, all optimized to run seamlessly on DGX hardware.

Q4: How does the DGX H100 scale for larger workloads? The DGX H100 is engineered for massive scalability. Using its high-speed ConnectX-7 NICs and InfiniBand fabric, multiple DGX H100 systems can be clustered together to form an NVIDIA DGX SuperPOD. This architecture allows organizations to build AI supercomputers with tens or hundreds of DGX systems, providing the necessary power to train foundational models with trillions of parameters.

Last update at December 2025