NVIDIA HGX B200 (8-GPU) Platform

Description

The HGX B200 (8-GPU) is a next-generation AI compute platform built on NVIDIA’s Blackwell architecture, optimized for training and deploying massive language models, generative AI, and scientific simulations. With 8 Blackwell SXM GPUs connected via NVLink 5.0 and NVSwitch, this platform delivers breakthrough performance, memory bandwidth, and scalability for data centers and AI superclusters.

Key Highlights

Architecture & GPU Configuration

• GPU Architecture: NVIDIA Blackwell
• GPU Count: 8 × Blackwell SXM
• Transformer Engine: Second-generation, optimized for FP8 and FP4 precision
• GPU Interconnect: NVLink 5.0 + NVSwitch
• GPU-to-GPU Bandwidth: 1.8 TB/s
• Total NVLink Bandwidth: 14.4 TB/s

Memory

• Total GPU Memory: 1.4 TB HBM3e
• Memory Type: HBM3e (High Bandwidth Memory)
• Memory Bandwidth per GPU: 600 GB/s
• Aggregate Bandwidth: 4.8 TB/s across all GPUs

Performance Metrics

These numbers reflect the platform’s ability to handle both training and inference at scale, with exceptional throughput for transformer-based models and mixed-precision workloads.

Connectivity & Expansion

• PCIe Generation: PCIe Gen5
• PCIe Bandwidth: 128 GB/s
• Networking Bandwidth: 0.8 TB/s
• Compatible with high-speed NICs and DPUs for cluster-scale deployments

Use Cases

• Training trillion-parameter LLMs
• Generative AI (text, image, video)
• Scientific computing and simulation
• Multi-modal inference at scale
• AI SuperPods and hyperscale data centers