NVIDIA H200 NVL Passive PCIe 141GB with NVIDIA AI Enterprise
NVIDIA H200 NVL Passive PCIe 141GB with NVIDIA AI Enterprise
This product is intended solely for businesses, government and educational institutions. Completing an order on the e-shop does not constitute a binding contract, but rather the submission of an enquiry.
The sale of this product is subject to export control, sanctions and other legal restrictions. Before an order can be accepted, the buyer, the end user, the country of delivery and the intended use must be verified. An enquiry can only be accepted after successful authorisation by the supplier.
If the required conditions are not met, the enquiry may be declined or the order cancelled without any entitlement to delivery of the goods.
If you are interested in purchasing an AI solution or accelerator, you can also email us at [email protected]. We will then prepare a proposal and an individual contract for you. For further information, please refer to our terms and conditions (Articles 3.33 and 3.34).
NVIDIA H200 NVL Passive PCIe 141GB with NVIDIA AI Enterprise graphics card
The NVIDIA H200 NVL Passive PCIe 141GB responds to this demand in a way that the competition is still catching up to: 141GB of HBM3e memory with a throughput of 4.8TB/s, NVIDIA Hopper architecture with a GH100 chip, and a performance of 3341 TFLOPS in FP8 and 1671 TFLOPS in FP16 for both inference and training of the largest AI models, directly in a PCIe rack without the need for specialised HGX infrastructure. Compared to the H100 NVL, the NVIDIA H200 NVL delivers up to 1.7× higher performance for LLM inference and 1.3× higher performance for HPC applications – and the included five-year NVIDIA AI Enterprise licence ensures production deployment from day one.
HBM3 / 141GB
Operating memory
PCIe 5.0 x16
Bus interface
None
Additional power supply
Tesla Hopper (GH100)
Chip architecture
1600W
Recommended source power
Key features of the NVIDIA H200 NVL Passive PCIe 141GB with NVIDIA AI Enterprise graphics card
- The card features 141GB of HBM3e memory with 4.8TB/s throughput on a 6144-bit bus – almost double the capacity and 1.4× more throughput than the H100 – for inference and training of the largest language models, scientific simulations, and HPC workloads without memory constraints
- NVIDIA Hopper architecture with a GH100 chip and a boost frequency of up to 1785MHz achieves 60 TFLOPS in FP32, 1671 TFLOPS in FP16/BF16, and 3341 TFLOPS in FP8 and INT8 for maximum AI inference performance and large-scale model training
- Multi-Instance GPU allows the card to be divided into up to 7 fully isolated instances of 16.5GB each, with their own memory, cache, and guaranteed quality of service for secure concurrent operation of multiple users or workloads on a single card
- Support for a 2- or 4-way NVIDIA NVLink bridge with 900GB/s throughput per GPU enables connecting up to four H200 NVL cards into a powerful multi-GPU cluster with larger shared memory for extremely large models and datasets
- A PCIe Gen 5 interface with 128GB/s throughput enables deployment in standard enterprise server enclosures without the need for specialised HGX or DGX infrastructure, significantly lowering the entry barrier for AI data centres
- Confidential Computing with hardware isolation protects sensitive data and AI models from unauthorised access, even in shared server infrastructure, to meet strict security and compliance requirements
- Seven NVDEC decoders and seven JPEG encoders provide hardware acceleration for video and image data processing for vision AI applications, media workflows, and large-scale data pipelines
- The card comes with a five-year NVIDIA AI Enterprise licence that includes NVIDIA NIM microservices, optimised frameworks for generative AI, computer vision, speech AI, and RAG systems with enterprise-grade security and support
- TBP of up to 600W (configurable) with a PCIe 5.0 x16 interface requires server infrastructure with adequate power and cooling for stable production operation
Memory throughput that changes AI inference
The biggest bottleneck in modern AI inference isn't computational power – it's moving data between memory and the processing cores. The NVIDIA H200 NVL PCIe 141GB radically removes this barrier. A total of 141GB of HBM3e with a throughput of 4.8TB/s ensures that the processing cores never wait for data. The H200 NVL can process models like Llama2 70B up to 1.7× faster than the H100 NVL, and GPT-3 175B 1.6× faster – all with an identical power profile, meaning the same power and cooling infrastructure requirements. For AI factories and supercomputing systems that need more performance without expanding their energy infrastructure, the NVIDIA H200 NVL is the natural choice.
PCIe form factor for data centres without HGX infrastructure
HGX-class performance in servers that previously couldn't accommodate it – that's exactly what the NVIDIA H200 NVL PCIe 141GB offers. While the SXM version requires a specialised HGX motherboard, the PCIe variant fits into standard enterprise servers like NVIDIA MGX and NVIDIA-Certified Systems with a capacity of up to eight GPUs in a single system. Support for a 2- or 4-way NVLink bridge with 900GB/s throughput per GPU then allows you to build a powerful multi-GPU cluster even without a proprietary interconnect system. For data centres that want to upgrade existing server infrastructure instead of completely replacing it, this is a strategic choice with an immediate return on investment.
5 years of NVIDIA AI Enterprise
Hardware without software is only half the solution. That's why the NVIDIA H200 NVL PCIe 141GB comes with a five-year NVIDIA AI Enterprise subscription included – a complete platform for developing and deploying generative AI, speech models, RAG systems, and agentic AI applications. NVIDIA NIM microservices enable the deployment of optimised AI models as containers with a lead time of minutes, while enterprise drivers and support ensure stability and security for mission-critical deployments. For organisations that need to move from AI experiments to a production infrastructure, the H200 NVL with AI Enterprise shortens the path from prototype to production to a minimum.
NVIDIA Hopper architecture
Defines the standards for modern AI inference and HPC computing
3958 TFLOPS in FP8
Hopper architecture AI inference performance for large-scale LLMs
141GB HBM3e
4.8TB/s throughput for inferencing and training the largest AI models
NVLink 900GB/s
Connect up to four GPUs into a multi-GPU cluster with extreme data throughput
7× MIG at 16.5GB each
Up to 7 isolated GPU instances for shared data centre environments
Confidential computing
Hardware-based data and model isolation to meet compliance and security requirements
PCIe Gen 5 form factor
Deployment in standard enterprise servers without HGX infrastructure
5 years of NVIDIA AI Enterprise
NIM microservices, enterprise support, and optimised AI frameworks includedSpecifications
Chip
Performance & usage
Memory
Frequency
Interface
Technology
Cooling system
Dimensions
Consumption
AI
NCNR (Non-Cancellable, Non-Returnable)
Tops
AI specifications
This product is not an independently functional unit and may require professional installation.
Specifications can be changed without notice. Images are for illustrative purposes only.
