Nvidia H200 NVL 141GB HBM3e PCIe 5.0 GPU - 900-21010-0040-000

The NVIDIA H200 NVL (900-21010-0040-000) is NVIDIA's flagship AI and HPC accelerator, featuring 141GB of HBM3e memory — the largest GPU memory capacity available. Built on the Hopper architecture with PCIe 5.0 x16 connectivity, it delivers 4.8TB/s memory bandwidth and exceptional FP8/FP16/BF16 throughput for large language model (LLM) inference, AI training, and high-performance computing workloads.

Key Features

• 141GB HBM3e — highest GPU memory capacity for large model inference
• 4.8TB/s memory bandwidth — 76% more than H100 SXM
• PCIe 5.0 x16 interface for broad server compatibility
• Hopper architecture — FP8, FP16, BF16, TF32, FP64 precision support
• NVLink bridge support for multi-GPU configurations
• Designed for LLM inference, AI training, and HPC workloads

Product Specifications

• Part Number: 900-21010-0040-000
• Manufacturer: NVIDIA
• OEM Status: Retail
• Model: H200 NVL
• Architecture: Hopper (GH100)
• Memory: 141GB HBM3e
• Interface: PCI-Express 5.0 x16
• Form Factor: SXM / PCIe (verify form factor for this SKU)

Processor Specifications

• Architecture: Hopper (GH100)
• CUDA Cores: 16,896
• Tensor Cores: 528 (4th Gen)
• NVLink: 4th Gen NVLink
• TDP: [Verify from NVIDIA datasheet]

Performance Specifications

• Memory Bandwidth: 4.8 TB/s
• FP8 Tensor: Up to 3,958 TFLOPS
• FP16 / BF16 Tensor: Up to 1,979 TFLOPS
• FP64 Tensor: Up to 67 TFLOPS
• Interface Bandwidth: PCIe 5.0 x16 (128 GB/s bidirectional)

Use Cases

• Large language model (LLM) inference — fits larger models in single GPU
• Generative AI training and fine-tuning
• High-performance computing (HPC) and scientific simulation
• Multi-GPU AI clusters via NVLink
• Data analytics and recommendation systems

Compatibility

• PCIe 5.0 x16 servers and workstations
• Compatible with PCIe 4.0 x16 slots (at reduced bandwidth)
• NVIDIA NVLink bridge for multi-GPU configurations
• Requires appropriate power delivery — verify server PSU capacity

Frequently Asked Questions

What is the difference between H200 NVL and H100?

The H200 NVL offers 141GB HBM3e vs H100's 80GB HBM2e/HBM3, with 4.8TB/s memory bandwidth vs 3.35TB/s. The larger memory capacity enables inference of significantly larger AI models without model sharding.

Is this an SXM or PCIe form factor?

The H200 NVL is available in both SXM and PCIe form factors. Verify the specific form factor for part number 900-21010-0040-000 from NVIDIA's official product page before ordering.

Does this require special cooling?

Yes, the H200 NVL requires active cooling. It is designed for data center servers with appropriate airflow — not suitable for standard desktop workstations.

The NVIDIA H200 NVL 900-21010-0040-000 delivers 141GB HBM3e with 4.8TB/s memory bandwidth on the Hopper architecture — the definitive GPU for large-scale AI inference, LLM training, and HPC workloads requiring maximum GPU memory capacity.