NVIDIA HGX B200 NVL8
The NVIDIA HGX B200 NVL8 is a high-density 8-GPU AI and high-performance computing platform built around eight NVIDIA Blackwell B200 Tensor Core GPUs connected through a high-bandwidth NVIDIA NVLink/NVSwitch fabric. With 1.44TB of aggregate HBM3e GPU memory, up to 64TB/s of aggregate GPU memory bandwidth, and up to 1.8TB/s GPU-to-GPU bandwidth, HGX B200 NVL8 is engineered for large-scale generative AI, deep learning, LLM training and inference, scientific computing, and enterprise AI infrastructure.
Description
The NVIDIA HGX B200 NVL8 is an enterprise-class accelerated computing platform designed to provide extreme GPU density and tightly coupled GPU-to-GPU communication for demanding artificial intelligence and high-performance computing workloads.
At the heart of the platform are eight NVIDIA B200 Blackwell GPUs in SXM form factor. Each B200 provides 180GB of HBM3e memory, resulting in a combined GPU memory capacity of approximately 1.44TB across the eight-GPU platform.
The eight GPUs provide up to 64TB/s of aggregate GPU memory bandwidth, enabling extremely high-throughput processing of large AI models, training datasets, scientific simulations, and other memory-intensive workloads.
A key feature of HGX B200 NVL8 is its high-speed GPU interconnect. The platform uses fifth-generation NVIDIA NVLink together with NVIDIA NVLink Switch technology, providing up to 1.8TB/s of GPU-to-GPU bandwidth and approximately 14.4TB/s of total NVLink bandwidth across the eight-GPU platform. This high-bandwidth fabric allows the GPUs to exchange model parameters, activations, gradients, and other data efficiently during distributed workloads.
The HGX B200 platform delivers up to 144 PFLOPS of FP4 Tensor Core AI performance and up to 72 PFLOPS of FP8 Tensor Core performance, with additional support for FP16/BF16, TF32, FP32, and FP64 workloads.
Unlike a conventional standalone graphics card, HGX B200 is an 8-GPU server platform/baseboard architecture. The exact CPU configuration, system memory, storage, networking adapters, chassis, power supplies, cooling solution, and management controller are determined by the OEM or NVIDIA-certified server manufacturer implementing the HGX B200 platform.
For networking, NVIDIA’s reference architecture supports high-speed 400Gb/s networking, including NVIDIA BlueField-3 SuperNICs and DPUs. The architecture is designed for scalable AI clusters, enabling HGX B200 systems to be interconnected through high-performance InfiniBand or Ethernet infrastructure.
HGX B200 is therefore particularly suited to organizations building AI factories, GPU clusters, NVIDIA-Certified systems, enterprise AI infrastructure, scientific computing platforms, and large-scale HPC environments.
Complete Technical Specification
| Specification | NVIDIA HGX B200 NVL8 |
|---|---|
| Product Type | 8-GPU Accelerated Computing Platform |
| Platform Family | NVIDIA HGX B200 |
| Architecture | NVIDIA Blackwell |
| GPU Configuration | 8 × NVIDIA B200 SXM GPUs |
| GPU Form Factor | SXM |
| GPU Memory per GPU | 180GB HBM3e |
| Total GPU Memory | 1.44TB HBM3e |
| GPU Memory Bandwidth per GPU | Up to 8TB/s |
| Aggregate GPU Memory Bandwidth | Up to 64TB/s |
| FP4 Tensor Performance | Up to 144 PFLOPS |
| FP8/FP6 Tensor Performance | Up to 72 PFLOPS |
| FP16/BF16 Tensor Performance | Up to 36 PFLOPS |
| TF32 Tensor Performance | Up to 18 PFLOPS |
| FP32 Performance | Up to 600 TFLOPS |
| GPU Interconnect | NVIDIA NVLink |
| NVLink Generation | 5th Generation |
| NVLink Switch | NVIDIA NVLink 5 Switch |
| GPU-to-GPU Bandwidth | Up to 1.8TB/s |
| Aggregate NVLink Bandwidth | Up to 14.4TB/s |
| CPU Platform | x86-based; OEM dependent |
| CPU Sockets | Minimum 2 recommended in NVIDIA reference architecture |
| CPU Core Recommendation | 56 physical cores per socket |
| System Memory | OEM/system dependent; NVIDIA reference architecture specifies minimum 1.5TB |
| Memory Bandwidth Requirement | Minimum 500GB/s system-memory bandwidth in reference architecture |
| PCIe | PCIe Gen5 |
| GPU PCIe Connectivity | 8 × Gen5 x16 links |
| DPU | NVIDIA BlueField-3 supported |
| SuperNICs | NVIDIA BlueField-3 SuperNIC |
| Networking | Up to 400Gb/s per adapter |
| Networking Protocols | InfiniBand / Ethernet, depending on configuration |
| Local Storage | OEM/system dependent |
| Management | BMC / platform management dependent on OEM implementation |
| Security | Secure Boot / TPM 2.0 supported in certified reference configurations |
| Cooling | High-performance server cooling; OEM dependent |
| Power | OEM/system dependent; each B200 GPU configurable up to 1kW |
| Server Form Factor | OEM dependent |
| Platform Type | Data-center / enterprise |
| Target Workloads | AI, ML, HPC, generative AI, LLMs, analytics |
Applications
Artificial Intelligence & Machine Learning
- Large Language Model training
- LLM inference
- Generative AI
- Transformer workloads
- Multimodal AI
- Computer vision
- Speech and language processing
- Recommendation systems
- AI model fine-tuning
High-Performance Computing
HGX B200 NVL8 provides the GPU density and interconnect performance required for:
- Scientific simulations
- Computational fluid dynamics
- Molecular modeling
- Weather and climate research
- Genomics
- Financial modeling
- Engineering simulations
- Physics and research workloads
Enterprise AI
Suitable for enterprise AI infrastructure requiring high GPU memory capacity and high-speed GPU-to-GPU communication.
AI Training Clusters
Multiple HGX B200 systems can be combined into larger GPU clusters using high-speed InfiniBand or Ethernet networking for distributed AI training and inference.
Data Analytics
The platform can accelerate large-scale analytics, machine learning pipelines, graph analytics, and data-intensive computational workloads.
Compatibility
The HGX B200 NVL8 is not a standalone PCIe graphics card. It is an 8-GPU server platform designed to be integrated by OEMs and NVIDIA-certified system manufacturers.
It is compatible with appropriately designed:
- NVIDIA HGX B200 server platforms
- NVIDIA-Certified systems
- x86 server CPU platforms
- NVIDIA BlueField-3 networking infrastructure
- 400Gb/s InfiniBand networks
- 400GbE Ethernet networks
- NVIDIA AI Enterprise
- CUDA
- NVIDIA NGC software ecosystem
- NVIDIA AI software stack
- High-performance NVMe storage
- Data-center rack and cooling infrastructure
Exact CPU, DIMM, storage, networking, chassis, PSU, and cooling compatibility depends on the specific OEM implementation.
Key Benefits
🚀 8-GPU Blackwell Architecture
Eight B200 GPUs provide a highly dense accelerated-computing platform for demanding AI and HPC workloads.
🧠 1.44TB HBM3e GPU Memory
The eight GPUs provide a combined 1.44TB of high-bandwidth HBM3e memory, enabling large models and data-intensive workloads to remain close to GPU compute resources.
⚡ Up to 64TB/s GPU Memory Bandwidth
The platform provides exceptional aggregate memory throughput for memory-intensive AI and HPC applications.
🔗 High-Speed NVLink Fabric
Fifth-generation NVLink and NVLink Switch technology provides up to 1.8TB/s GPU-to-GPU bandwidth, helping minimize communication bottlenecks between GPUs.
📊 144 PFLOPS FP4 AI Performance
The eight-GPU platform can deliver up to 144 PFLOPS of FP4 Tensor Core performance, targeting modern generative AI and accelerated inference workloads.
🌐 400Gb/s-Class Networking
NVIDIA reference architectures support high-speed 400Gb/s networking for scalable multi-node AI clusters.
🏢 Enterprise-Scale Infrastructure
HGX B200 is designed for professional data centers, AI factories, HPC environments, cloud infrastructure, and large-scale research systems.
🔧 OEM Flexibility
Because HGX B200 is a platform rather than a fixed server, system manufacturers can configure CPU, memory, storage, networking, cooling, and chassis components for specific deployment requirements.
SEO Title
NVIDIA HGX B200 NVL8 – 8× B200 Blackwell GPUs, 1.44TB HBM3e, 64TB/s AI Platform



