NVIDIA H200 NVL, Passive PCIe, Blade-Server - 141 GB HBM3

The NVIDIA H200 NVL is a data centre GPU designed for generative AI, LLM inference, and high performance computing. With 141 GB of HBM3e and a memory bandwidth of 4,800 GB/s, it processes large models and extensive datasets directly in fast GPU memory. 16,896 CUDA Cores and 528 Tensor Cores provide the computing power for highly parallelised AI and HPC workloads.

• 141 GB HBM3e with ECC for large models and memory-intensive datasets
• 4,800 GB/s memory bandwidth for high data throughput in AI and HPC workloads
• 16,896 CUDA Cores and 528 Tensor Cores for massively parallel computations
• PCIe 5.0 x16 and NVLink for scalable multi-GPU configurations
• Passive dual-slot cooling for use in specially designed GPU servers

141 GB HBM3e for large models and datasets

For large language models and memory-intensive HPC applications, the available GPU memory capacity determines how much data can be processed directly. The H200 NVL provides 141 GB of HBM3e with ECC for this purpose. A memory bandwidth of 4,800 GB/s ensures that the compute units are supplied with data quickly. This is particularly relevant for LLM inference, generative AI, and data-intensive simulations.

Computing power for AI and high performance computing

The Hopper-based GPU combines 16,896 CUDA Cores with 528 Tensor Cores. This provides specialised compute resources for matrix operations and highly parallelised workloads. Supported formats include FP64, FP32, TF32, BF16, FP16, and FP8. This allows you to use the H200 NVL for both scientific calculations and AI models with varying precision requirements.

Designed for scalable GPU servers

The H200 NVL connects to the host system via PCIe 5.0 x16. For multi-GPU systems, it supports configurations with NVIDIA NVLink and up to 900 GB/s connection bandwidth. Up to seven MIG instances, each with 16.5 GB, also enable GPU resources to be allocated to separate workloads. The passive dual-slot cooling requires a server designed for this purpose with sufficient airflow, making it particularly suitable for professional data centre environments.