The NVIDIA H200 NVL 141 GB (AI) is a datacenter GPU accelerator based on the Hopper architecture, designed for the most demanding workloads in generative AI (LLM), large-scale inference, HPC, and analytics. The NVL version is designed for PCIe format servers, with high memory capacity and high-speed NVLink interconnection for multi-GPU deployments.
Highlights
-
Very large HBM3e memory (141 GB): ideal for large AI models, large contexts, large batches, and heavy datasets.
-
Massive memory bandwidth: excellent for "memory-bound" workloads (LLM, HPC).
-
AI optimized: Tensor Cores + Transformer Engine to accelerate modern calculations (FP8/FP16/BF16/TF32 depending on use).
-
Designed for 24/7 server use: PCIe integration, datacenter chassis-oriented cooling, sustained performance.
Technical Details (Key)
-
Architecture: NVIDIA Hopper (H200)
-
Memory: 141 GB HBM3e
-
Memory Bandwidth: 4.8 TB/s
-
Interconnect:
-
Power (TDP / TGP / profile): 600 W to 700 W depending on profile and server integration
-
MIG (Multi-Instance GPU): up to 7 instances (partitioning for multi-tenancy)
-
Hardware decoding: 7× NVDEC + 7× JPEG (useful for video/vision pipelines)
-
Format: PCIe dual-slot, generally air-cooled / server-adapted cooling
For Whom?
-
AI / LLM (inference and training): large-scale serving, large models, large contexts.
-
HPC / simulation: intensive computing with high memory pressure.
-
Datacenters and clouds: dense multi-GPU deployments, virtualization/partitioning via MIG.
