NVIDIA B200 SXM 180GB – Blackwell Data Center GPU

Description

The NVIDIA B200 SXM 180GB is part of NVIDIA’s Blackwell GPU architecture, developed specifically for next-generation AI and accelerated data-center computing. It combines extremely high compute throughput with a large high-bandwidth memory subsystem, making it suitable for workloads where both processing performance and memory capacity are critical.

The production B200 SXM provides 180GB of HBM3e memory with up to 7.7TB/s of memory bandwidth. This large and extremely fast memory pool enables efficient processing of large language models (LLMs), transformer-based workloads, scientific simulations, recommendation systems, and other data-intensive applications.

The GPU provides up to 1,000W configurable thermal design power, reflecting its role as a high-density data-center accelerator. It uses the SXM6 form factor and is intended for integration into systems such as NVIDIA HGX B200 platforms, where multiple B200 GPUs are connected through a high-speed NVLink fabric.

For AI workloads, the B200 introduces Blackwell’s fifth-generation Tensor Core architecture and the second-generation Transformer Engine, supporting modern reduced-precision formats including FP4, FP6, FP8, BF16, and FP16. The accelerator can deliver up to 18 PFLOPS of sparse FP4 Tensor Core performance, making it particularly well suited to generative AI inference and training workloads.

The B200 also supports fifth-generation NVLink, providing up to 1.8TB/s of GPU interconnect bandwidth for high-speed communication between GPUs in supported multi-GPU platforms.

3. Complete Technical Specification

* Tensor Core figures depend on precision mode and whether structured sparsity is utilized.

NVIDIA’s current HGX documentation specifies 180GB HBM3e per B200 GPU, 1.44TB across an 8-GPU HGX B200 system, and up to 8TB/s-class bandwidth at the platform level, while NVIDIA’s B200-specific product documentation lists the B200 GPU at up to 7.7TB/s memory bandwidth.

4. Applications

The NVIDIA B200 SXM is designed for demanding accelerated-computing applications including:

• Generative AI training and inference
• Large Language Models (LLMs)
• Transformer-based AI workloads
• AI model fine-tuning
• Deep learning and machine learning
• Large-scale inference services
• Scientific and technical computing
• High-Performance Computing (HPC)
• Data analytics
• Recommendation systems
• Computer vision
• Natural language processing
• Simulation and computational science
• AI research and development
• Enterprise AI infrastructure
• Cloud GPU infrastructure
• Multi-GPU AI clusters

NVIDIA’s own performance resources list B200 configurations in large-scale training and inference workloads, including 8-GPU DGX B200 systems.

5. Compatibility

The B200 SXM 180GB is not a standard PCIe graphics card. It is an SXM6 accelerator designed for specialized data-center platforms.

Compatible Platform Types

• NVIDIA HGX B200 systems
• NVIDIA-Certified B200 server platforms
• Qualified OEM B200 servers
• Multi-GPU AI/HPC servers designed for SXM6 modules
• Specialized air-cooled or liquid-cooled B200 platforms

The NVIDIA HGX B200 platform uses eight B200 GPUs, providing 1.44TB of total HBM3e GPU memory and up to 64TB/s aggregate GPU memory bandwidth according to NVIDIA’s current enterprise architecture documentation.

Important Compatibility Note

The B200 SXM requires a compatible SXM6 baseboard, power delivery system, thermal solution, firmware, and system architecture. It should not be purchased as a drop-in replacement for PCIe GPUs such as the H100 PCIe or H200 PCIe.

6. Key Benefits

• 180GB HBM3e memory for extremely large AI models and datasets
• Up to 7.7TB/s memory bandwidth
• NVIDIA Blackwell architecture
• Fifth-generation Tensor Cores
• Second-generation Transformer Engine
• Support for FP4, FP6, FP8, BF16, FP16 and TF32
• Up to 18 PFLOPS sparse FP4 Tensor performance
• Up to 1.8TB/s fifth-generation NVLink bandwidth
• Designed for high-density multi-GPU computing
• Supports Multi-Instance GPU (MIG) configurations
• Optimized for generative AI and LLM workloads
• Suitable for large-scale HPC and scientific computing
• Designed for NVIDIA HGX B200 infrastructure
• Enables high-bandwidth GPU-to-GPU communication
• Enterprise-class accelerator architecture
• Available in qualified air- and liquid-cooled server configurations

7. SEO Title

NVIDIA B200 SXM 180GB HBM3e Blackwell Tensor Core GPU – 7.7TB/s AI & HPC Accelerator