NVIDIA H100NVL

NVIDIA Graphics Card

The NVIDIA H100 NVL is a Hopper-architecture data centre GPU designed specifically for large language model inference, carrying 94GB of HBM3 high-bandwidth memory per card with 3.9 TB/s of memory bandwidth. A built-in Transformer Engine with FP8 precision delivers several times the large-model inference throughput of the previous generation. The PCIe 5.0 dual-slot passively cooled design can be paired through an NVLink bridge for 600 GB/s of card-to-card bandwidth, and Multi-Instance GPU (MIG) allows partitioning into up to seven independent instances. SWIT Technology supplies the H100 NVL as part of complete 4U and 6U GPU server builds.

Specifications

ArchitectureNVIDIA Hopper architecture (GH100)
4th-generation Tensor Cores
Built-in Transformer Engine with FP8 support
GPU Memory94GB HBM3 high-bandwidth memory
3.9 TB/s memory bandwidth
System InterfacePCIe 5.0 x16
InterconnectNVLink bridge support, 600 GB/s card-to-card bandwidth
VirtualisationMulti-Instance GPU (MIG), up to 7 independent instances
NVIDIA vGPU support
Power350W to 400W max power consumption (configurable)
Cooling & Form FactorPassive cooling (chassis airflow)
Dual-slot full-height full-length, for rack servers
Software SupportCUDA, TensorRT-LLM, NVIDIA AI Enterprise, NCCL
Use CasesLarge language model inference, generative AI services, deep learning training, HPC
載入中…