English 
| Product Name: | |
|---|---|
| Platform: | |
| GPU Architecture: | |
| GPU Type: | |
| GPU Interconnect: | |
| Memory Type: | |
| Memory Capacity: | |
| Quantity: | |
H100 Server
NVIDIA
The NVIDIA HGX H100 and H200 GPU servers represent Hopper-architecture computing platforms built for large-scale AI training, inference, and high-performance computing workloads. Deployed as 4-GPU or 8-GPU server configurations via the NVIDIA HGX baseboard, these systems support distributed training clusters and enterprise data center expansion. All specifications below reference NVIDIA official product datasheets.
Both H100 and H200 GPUs are built on the NVIDIA Hopper architecture with fourth-generation Tensor Cores. Each GPU delivers 4 PFLOPS of FP8 tensor throughput, supporting Transformer Engine acceleration for large language model training. The H100 SXM variant operates at up to 700W TDP with configurable power profiles.
Parameter | H100 SXM | H200 SXM |
Memory Capacity | 80 GB HBM3 | 141 GB HBM3e |
Memory Bandwidth | 3.35 TB/s | 4.8 TB/s |
TDP (Configurable) | Up to 700W | Up to 700W |
Form Factor | SXM | SXM |
The H200 variant provides 1.4× higher memory bandwidth and nearly double the memory capacity compared with the H100 SXM configuration, supporting larger batch sizes and extended context windows for LLM workloads.
HGX H100 and H200 servers use fourth-generation NVLink for GPU-to-GPU communication at 900 GB/s per GPU. Integrated NVSwitch technology enables all-to-all connectivity across 8 GPUs within a single node, reducing communication bottlenecks in distributed training pipelines.
Each GPU supports PCIe Gen5 ×16 connectivity at 128 GB/s for host communication. Systems can be integrated with NDR 400Gb/s InfiniBand or 400GbE networking for multi-node cluster deployment, compatible with NVIDIA Magnum IO software stack.
With high-bandwidth memory and NVLink scaling, 8-GPU HGX H200 configurations support training and fine-tuning of transformer models with tens of billions of parameters. Higher memory capacity reduces the need for model parallelism overhead in medium-scale LLM projects.
FP64 and TF32 compute capabilities support computational fluid dynamics, molecular dynamics, and weather simulation workloads. Multi-Instance GPU (MIG) partitioning allows up to seven independent instances per GPU for shared tenancy environments.
HGX H100 and H200 servers are available in 8U rackmount form factors from NVIDIA-Certified Systems partners. 8-GPU configurations require adequate liquid or advanced air cooling infrastructure and redundant high-wattage power supplies.
Systems support NVIDIA CUDA 12+, cuDNN, TensorRT, and all major AI frameworks including PyTorch, TensorFlow, and JAX. NVIDIA AI Enterprise provides enterprise-grade support and security patches for commercial procurement environments.
The H200 shares the same Hopper compute architecture as the H100 but upgrades to HBM3e memory with 141 GB capacity and 4.8 TB/s bandwidth per GPU. The H100 uses 80 GB HBM3 at 3.35 TB/s. Both deliver 4 PFLOPS of FP8 tensor performance per GPU.
The NVIDIA HGX platform supports both 4-GPU and 8-GPU baseboard configurations for H100 and H200 variants. 8-GPU configurations include NVSwitch for full-mesh NVLink connectivity.
Servers are available through NVIDIA-Certified Systems partners in standard rackmount form factors. They require compatible power distribution and cooling infrastructure, and support standard data center networking via InfiniBand or Ethernet adapters.
