NVIDIA DGX H100 AI Supercomputer
The NVIDIA DGX H100 is an enterprise-class AI supercomputer designed for large-scale artificial intelligence, generative AI, deep learning, high-performance computing (HPC), data analytics, and large language model workloads.
Powered by 8 NVIDIA H100 Tensor Core GPUs, DGX H100 provides 640GB of total GPU memory, connected through NVIDIA’s fourth-generation NVLink and NVSwitch architecture. The system delivers up to 32 PFLOPS of FP8 AI performance and provides high-bandwidth networking through NVIDIA ConnectX-7 adapters.
The system combines GPU acceleration, dual Intel Xeon processors, 2TB of system memory, high-speed NVMe storage, ConnectX-7 networking, NVSwitch fabric, and NVIDIA’s optimized DGX software stack into a fully integrated AI infrastructure platform.
Description
The NVIDIA DGX H100 is the fourth-generation NVIDIA DGX system and was specifically engineered to provide a complete infrastructure platform for enterprise-scale AI.
Rather than requiring organizations to integrate GPUs, networking, storage, and software independently, DGX H100 combines these components into a validated system optimized for large distributed AI workloads.
Eight NVIDIA H100 Tensor Core GPUs
At the heart of DGX H100 are 8 NVIDIA H100 Tensor Core GPUs based on the NVIDIA Hopper architecture.
Together, the eight GPUs provide:
- 640GB total HBM3 GPU memory
- 80GB HBM3 per GPU
- Up to 3.35TB/s memory bandwidth per GPU
- 8,192-bit aggregate GPU memory interface
- Fourth-generation Tensor Cores
- Fourth-generation NVLink connectivity
The H100 GPU is optimized for transformer-based AI, generative AI, scientific computing, deep learning, recommendation systems, and other computationally intensive workloads. NVIDIA specifies up to 34 TFLOPS FP64, 67 TFLOPS FP32, and 3,958 TFLOPS FP8 Tensor Core performance with sparsity for the H100 SXM configuration.
NVIDIA NVLink and NVSwitch
The eight GPUs are interconnected using NVIDIA’s high-speed GPU fabric.
DGX H100 incorporates:
- 4 × fourth-generation NVIDIA NVSwitch
- 18 NVLink connections per GPU
- Up to 900GB/s bidirectional GPU-to-GPU bandwidth per GPU
- Up to 7.2TB/s aggregate bidirectional GPU-to-GPU bandwidth
This architecture allows the eight H100 GPUs to operate as a tightly interconnected accelerated computing platform rather than as independent graphics processors.
High-Performance Networking
DGX H100 integrates NVIDIA ConnectX-7 networking for high-speed cluster communication.
The system provides:
- 8 × single-port ConnectX-7 400Gb/s InfiniBand adapters
- 2 × dual-port ConnectX-7 Ethernet adapters
- Up to 400Gb/s InfiniBand
- Up to 400GbE Ethernet
- Support for 200GbE, 100GbE, 50GbE, 40GbE, 25GbE and 10GbE depending on interface/configuration
The networking architecture provides up to approximately 1TB/s of peak bidirectional network bandwidth across the system.
This makes DGX H100 suitable for multi-node AI clusters, DGX SuperPOD environments, distributed training, large-scale inference, and high-performance storage.
Dual Intel Xeon Platinum Processors
DGX H100 uses two Intel Xeon Platinum 8480C processors.
Together they provide:
- 112 CPU cores total
- 56 cores per processor
- Base frequency: 2.0GHz
- All-core turbo: up to 2.9GHz
- Maximum turbo: up to 3.8GHz
The CPU subsystem handles system management, data preparation, orchestration, storage, networking, and host-side application workloads.
2TB System Memory
The system includes 2TB of system DDR5 memory using 32 DIMMs.
This large CPU memory capacity supports:
- Large datasets
- Data preprocessing
- AI training pipelines
- Multi-GPU workloads
- HPC applications
- High-performance databases
- Large-scale analytics
High-Speed NVMe Storage
DGX H100 includes approximately 30TB of internal NVMe storage in the standard configuration.
The storage architecture consists of:
- 2 × 1.92TB NVMe M.2 SSDs configured as RAID 1 for the operating system
- 8 × 3.84TB NVMe U.2 self-encrypting drives configured as RAID 0 for data cache
This provides fast local storage for operating-system files, datasets, model checkpoints, temporary data, and AI workloads.
BlueField-3 and Infrastructure Acceleration
DGX H100 incorporates NVIDIA networking and infrastructure acceleration technologies designed to offload networking, storage, and security services.
NVIDIA’s original DGX H100 announcement also identifies two NVIDIA BlueField-3 DPUs within the system architecture for infrastructure offload and isolation.
Enterprise Software Stack
DGX H100 is designed to operate with NVIDIA’s optimized DGX software environment.
The platform supports:
- NVIDIA DGX OS
- CUDA
- NVIDIA GPU drivers
- NVIDIA Base Command
- NVIDIA AI Enterprise
- NVIDIA Magnum IO
- NVIDIA networking software
- GPU Direct Storage
- AI and HPC frameworks
Current NVIDIA documentation lists DGX H100 as supported by DGX OS 8, while DGX OS 7 also supports the system.
3. Complete Technical Specification
| Specification | Details |
|---|---|
| Product Name | NVIDIA DGX H100 |
| Product Type | Enterprise AI Supercomputer / AI Server |
| GPU Architecture | NVIDIA Hopper |
| GPU | 8 × NVIDIA H100 Tensor Core GPU |
| GPU Form Factor | SXM |
| GPU Memory | 640GB total |
| Memory per GPU | 80GB HBM3 |
| GPU Memory Bandwidth | Up to 3.35TB/s per GPU |
| GPU-to-GPU Interconnect | 4th-generation NVIDIA NVLink |
| NVLink per GPU | 18 connections |
| GPU-to-GPU Bandwidth | Up to 900GB/s bidirectional per GPU |
| NVSwitch | 4 × NVIDIA NVSwitch |
| Aggregate GPU Fabric Bandwidth | Up to 7.2TB/s bidirectional |
| FP8 AI Performance | Up to 32 PFLOPS with sparsity |
| CPU | 2 × Intel Xeon Platinum 8480C |
| CPU Cores | 112 total / 56 per CPU |
| CPU Base Frequency | 2.0GHz |
| CPU All-Core Turbo | Up to 2.9GHz |
| CPU Maximum Turbo | Up to 3.8GHz |
| System Memory | 2TB |
| Memory Configuration | 32 × DIMMs |
| Cluster Networking | 8 × ConnectX-7 single-port 400Gb/s adapters |
| Cluster Network Interface | OSFP |
| InfiniBand | Up to 400Gb/s |
| Ethernet Networking | Up to 400GbE |
| Ethernet Speeds | 400/200/100/50/40/25/10GbE |
| Storage Networking | 2 × ConnectX-7 dual-port Ethernet adapters |
| OS Storage | 2 × 1.92TB NVMe M.2 |
| OS RAID | RAID 1 |
| Data Cache Storage | 8 × 3.84TB NVMe U.2 SED |
| Data RAID | RAID 0 |
| Total Internal NVMe Capacity | Approximately 30TB |
| BMC | Integrated |
| BMC Network | 1GbE RJ45 |
| Management Protocols | Redfish, IPMI, SNMP, KVM, Web UI |
| Power Supplies | 6 × 3.3kW |
| Maximum System Power | Approximately 10.2kW |
| Operating Temperature | 5°C–30°C |
| System Architecture | x86-64 |
| Operating System | NVIDIA DGX OS |
| Primary Software | CUDA, NVIDIA AI Enterprise, Base Command, Magnum IO |
| Primary Workloads | AI, Generative AI, LLMs, HPC, analytics |
| System Class | Enterprise data-center infrastructure |
The current NVIDIA user guide confirms the 8-H100/640GB configuration, dual Xeon 8480C CPUs, 2TB system memory, 30TB-class NVMe storage, ConnectX-7 networking, four NVSwitches, and six 3.3kW power supplies.
4. Applications
Artificial Intelligence & Machine Learning
- Large language model training
- Generative AI
- Transformer models
- Deep learning
- Recommendation systems
- Computer vision
- Speech and language AI
- Multimodal AI
- Large-scale inference
High-Performance Computing
- Scientific simulations
- Climate modeling
- Molecular dynamics
- Computational chemistry
- Engineering simulation
- Physics research
- Financial modeling
- Genomics and life sciences
Enterprise AI
- Enterprise AI platforms
- AI-powered analytics
- Intelligent automation
- Natural language processing
- Customer intelligence
- Recommendation engines
- Predictive analytics
AI Infrastructure
- NVIDIA DGX SuperPOD
- AI training clusters
- Multi-node GPU infrastructure
- Private AI clouds
- Enterprise AI factories
- Research supercomputing
5. Compatibility
The DGX H100 is a complete integrated data-center system, rather than a conventional server intended for component-level GPU upgrades.
It is designed for deployment with:
- NVIDIA DGX SuperPOD infrastructure
- NVIDIA DGX BasePOD architectures
- InfiniBand networks
- 400GbE networks
- NVIDIA Quantum networking
- NVIDIA ConnectX-7 networking
- High-performance NVMe storage
- NVIDIA AI Enterprise
- NVIDIA Base Command
- CUDA-based applications
- Linux AI/HPC environments
- Enterprise AI clusters
For cluster deployment, the system is designed to connect through its high-speed ConnectX-7 networking to other DGX systems and compatible NVIDIA networking infrastructure.
Important: DGX H100 should not be confused with a standard 8-GPU HGX H100 server. DGX H100 is an NVIDIA-integrated system with validated GPU, NVSwitch, networking, storage, CPU, software, and management components.
6. Key Benefits
- 8 × NVIDIA H100 Tensor Core GPUs
- 640GB total HBM3 GPU memory
- Up to 32 PFLOPS FP8 AI performance
- High-bandwidth 4th-generation NVLink
- 4 × NVIDIA NVSwitch
- Up to 7.2TB/s aggregate GPU fabric bandwidth
- 112 Intel Xeon CPU cores
- 2TB system memory
- Approximately 30TB NVMe storage
- 400Gb/s ConnectX-7 networking
- InfiniBand and Ethernet connectivity
- Integrated BMC and enterprise management
- Designed for NVIDIA DGX SuperPOD environments
- Optimized for generative AI and large language models
- Enterprise-grade AI infrastructure
- Integrated NVIDIA software ecosystem
- High-speed GPU, network, and storage data paths
- Suitable for demanding AI, HPC and analytics workloads
7. SEO Title
NVIDIA DGX H100 AI Supercomputer – 8× H100 640GB, 2TB RAM, 400Gb/s Networking
Related products
-
Cisco CISCO1941/K9 1941 Integrated Services Router
Read more
Cisco CISCO1941/K9 1941 Integrated Services Router is a modular enterprise router designed for secure and reliable branch-office and small-business networking. It features 2 Gigabit Ethernet interfaces, 2 EHWIC expansion slots, 256 MB CompactFlash storage, 512 MB DRAM, and IP Base software, providing flexible connectivity and expansion for professional network deployments.
Quick View -
QLogic 7023303 Dual-Port 16Gb Fibre Channel / 10Gb FCoE Host Bus Adapter
Read more QLogic 7023303 Dual-Port 16Gb Fibre Channel / 10Gb FCoE Host Bus Adapter
The QLogic 7023303 is a high-performance dual-port Host Bus Adapter designed for enterprise SAN and data-center environments. It delivers up to 16Gbps Fibre Channel or 10Gbps Fibre Channel over Ethernet (FCoE) connectivity through two SFP+ ports and a PCI Express x8 host interface, providing flexible, high-bandwidth server-to-storage connectivity.
Quick View -
Dell PowerEdge R660
Read more
The Dell PowerEdge R660 is a high-performance 1U dual-socket enterprise rack server designed for high-density virtualization, databases, analytics, mixed workloads, and demanding data-center applications. It supports up to two 4th Generation Intel Xeon Scalable processors with up to 56 cores per processor, 32 DDR5 DIMM slots with up to 8TB RDIMM memory, high-speed SAS/SATA/NVMe storage, PCIe Gen 4 and Gen 5 expansion, optional accelerators, and iDRAC9 remote management. Current Dell documentation also supports configurations with 5th Generation Intel Xeon Scalable processors with up to 64 cores per processor.
Quick View -
Samsung-960GB Enterprise SSD
Read more
The Samsung 960GB Enterprise SSD is a reliable solid-state storage solution designed for enterprise servers, data centers, workstations, and professional applications. Offering 960GB of storage capacity, it provides fast data access, low latency, and dependable performance for virtualization, databases, application hosting, and other demanding workloads.
Quick View
