NVIDIA DGX B200 AI Supercomputer

Description

The NVIDIA DGX B200 represents a major generation of NVIDIA’s integrated DGX AI infrastructure, combining the Blackwell architecture with high-bandwidth GPU interconnects, enterprise networking, large system memory, high-speed storage, and a comprehensive NVIDIA software stack.

It is designed to operate as an integrated AI platform rather than simply an 8-GPU server, providing the infrastructure required for large-scale model training, inference, fine-tuning, and data-intensive accelerated computing.

Eight NVIDIA Blackwell B200 GPUs

DGX B200 incorporates 8 NVIDIA B200 Tensor Core GPUs.

The eight GPUs provide:

1.44TB total HBM3e memory
180GB HBM3e per GPU
Up to 64TB/s aggregate GPU memory bandwidth
Fifth-generation Tensor Cores
Second-generation Transformer Engine
Fourth-generation NVLink
Advanced FP4, FP6, FP8, BF16, FP16 and TF32 acceleration

The B200 GPUs are designed specifically for modern AI workloads, particularly transformer-based models and generative AI.

Blackwell Architecture

NVIDIA Blackwell introduces major advances in AI computing, including:

Fifth-generation Tensor Cores
Second-generation Transformer Engine
FP4 precision support
FP6 precision support
Advanced sparsity
Decompression Engine
Confidential Computing capabilities
High-speed GPU interconnects

These technologies allow DGX B200 to accelerate both training and inference while improving the efficiency of large-scale AI workloads.

1.44TB of HBM3e GPU Memory

Each B200 GPU provides 180GB of HBM3e memory.

Across eight GPUs, DGX B200 provides:

8 × 180GB = 1.44TB HBM3e

The memory subsystem delivers approximately 64TB/s aggregate GPU memory bandwidth, enabling extremely high-throughput movement of model parameters, activations, datasets, and intermediate results.

This large shared GPU-memory capacity is particularly important for large language models and other AI workloads with very large parameter sets.

NVIDIA NVLink and NVLink Switch

DGX B200 uses NVIDIA’s high-speed GPU interconnect architecture to allow the eight GPUs to communicate efficiently.

The system incorporates:

Fourth-generation NVLink
Fifth-generation NVLink Switch technology
High-bandwidth GPU-to-GPU communication
Low-latency communication between GPUs

This enables DGX B200 to function as a tightly integrated multi-GPU computing platform for distributed AI workloads.

AI Performance

DGX B200 provides up to approximately:

72 PFLOPS FP8 AI performance with sparsity
36 PFLOPS FP16/BF16 AI performance with sparsity
18 PFLOPS TF32 performance with sparsity

The exact performance depends on precision, workload, sparsity, and whether the benchmark represents theoretical peak or application performance.

NVIDIA positions DGX B200 specifically for training and inference of large AI models, including generative AI and trillion-parameter-class workloads.

Dual Intel Xeon CPUs

DGX B200 incorporates two Intel Xeon Platinum 8570 processors.

Together, the CPUs provide:

112 CPU cores
56 cores per processor
High-throughput host processing
Large PCIe connectivity
Support for large system-memory configurations

The CPU subsystem handles host-side application processing, data preparation, orchestration, storage, networking, and system management.

2TB System Memory

The DGX B200 includes 2TB of DDR5 system memory.

This provides substantial host-memory capacity for:

Large datasets
Data preprocessing
AI pipelines
Model staging
HPC workloads
Large-scale analytics
Multi-GPU applications

30TB NVMe Storage

DGX B200 includes approximately 30TB of internal NVMe storage.

The standard configuration uses:

2 × 1.92TB NVMe M.2 drives for the operating system
8 × 3.84TB NVMe U.2 drives for data/cache storage

The high-speed NVMe subsystem provides local storage for AI models, datasets, checkpoints, containers, temporary data, and operating-system resources.

High-Speed ConnectX-7 Networking

DGX B200 integrates NVIDIA ConnectX-7 network adapters for high-speed external connectivity.

The system provides:

8 × ConnectX-7 single-port 400Gb/s adapters
2 × ConnectX-7 dual-port 200Gb/s adapters
Up to 1.6Tb/s aggregate network bandwidth
InfiniBand and Ethernet support

This high-speed networking architecture is designed for multi-node DGX clusters, AI factories, DGX SuperPOD environments, distributed training, and high-performance storage.

BlueField-3 DPU

DGX B200 incorporates NVIDIA BlueField-3 DPUs for infrastructure acceleration and isolation.

These DPUs can offload infrastructure services such as:

Networking
Storage
Security
Virtualization
Infrastructure management

This allows host CPUs and GPUs to remain focused on application and AI workloads.

Enterprise AI Software

DGX B200 is designed around NVIDIA’s integrated software ecosystem.

Supported software technologies include:

NVIDIA DGX OS
CUDA
NVIDIA AI Enterprise
NVIDIA Base Command
NVIDIA Magnum IO
NVIDIA NIM
TensorRT
TensorRT-LLM
PyTorch
NVIDIA GPU drivers
NVIDIA networking software

This provides a validated environment for developing, training, deploying, and managing enterprise AI workloads.

3. Complete Technical Specification

Specification	Details
Product Name	NVIDIA DGX B200
Product Type	Enterprise AI Supercomputer / AI Server
GPU Architecture	NVIDIA Blackwell
GPU	8 × NVIDIA B200 Tensor Core GPU
GPU Form Factor	SXM
GPU Memory per GPU	180GB HBM3e
Total GPU Memory	1.44TB HBM3e
Aggregate GPU Memory Bandwidth	Up to 64TB/s
Tensor Cores	5th Generation
Transformer Engine	2nd Generation
RT Cores	4th Generation
FP8 Performance	Up to 72 PFLOPS with sparsity
FP16/BF16 Performance	Up to 36 PFLOPS with sparsity
TF32 Performance	Up to 18 PFLOPS with sparsity
GPU Interconnect	4th-generation NVLink
NVLink Switch	5th-generation NVLink Switch
CPU	2 × Intel Xeon Platinum 8570
CPU Cores	112 total
System Memory	2TB DDR5
High-Speed Networking	ConnectX-7
InfiniBand Networking	Up to 400Gb/s
Ethernet Networking	Up to 400GbE
Aggregate Network Bandwidth	Up to 1.6Tb/s
High-Speed Network Adapters	8 × single-port 400Gb/s ConnectX-7
Additional Network Adapters	2 × dual-port 200Gb/s ConnectX-7
DPU	NVIDIA BlueField-3
Internal Storage	Approximately 30TB NVMe
OS Storage	2 × 1.92TB NVMe M.2
Data Storage	8 × 3.84TB NVMe U.2
Operating System	NVIDIA DGX OS
AI Software	CUDA, NVIDIA AI Enterprise, Base Command, Magnum IO
Form Factor	10U data-center system
Primary Workloads	AI training, inference, generative AI, HPC
Deployment	Enterprise data center / AI factory

4. Applications

Artificial Intelligence & Machine Learning

Large language model training
Generative AI
Transformer models
Multimodal AI
Computer vision
Speech and language models
Recommendation systems
AI inference
Model fine-tuning

High-Performance Computing

Scientific simulations
Engineering simulation
Computational chemistry
Molecular dynamics
Climate modeling
Genomics
Financial modeling
Large-scale numerical computing

Enterprise Generative AI

Enterprise LLM platforms
Retrieval-augmented generation (RAG)
AI agents
Natural-language processing
Enterprise knowledge assistants
Generative image and video workloads
AI-powered analytics

AI Infrastructure

DGX SuperPOD
NVIDIA AI factories
Private AI clouds
Large GPU clusters
Multi-node distributed training
Enterprise AI infrastructure
Research supercomputing

5. Compatibility

DGX B200 is a complete NVIDIA-integrated data-center system and is not intended to be treated like a conventional workstation where individual GPUs are freely substituted.

It is designed for integration with:

NVIDIA DGX SuperPOD infrastructure
NVIDIA AI factories
NVIDIA Quantum InfiniBand networking
NVIDIA Spectrum Ethernet networking
ConnectX-7 adapters
BlueField-3 DPUs
High-performance NVMe storage
NVIDIA AI Enterprise
NVIDIA Base Command
CUDA-based AI applications
PyTorch and TensorRT workloads
Multi-node GPU clusters

For cluster deployments, the system can use high-speed InfiniBand or Ethernet networking to communicate with additional DGX B200 systems and compatible NVIDIA networking infrastructure.

Compatibility Note: DGX B200 uses an integrated NVIDIA architecture with SXM GPUs, NVLink/NVSwitch fabric, specialized cooling, networking, storage, and power infrastructure. Individual components should only be replaced with NVIDIA-approved or system-qualified equivalents.