H100 NVL is designed to scale support of Large Language Models in mainstream PCIe-based server systems. With increased raw performance, bigger, faster HBM3 memory and NVLink connectivity via bridges, mainstream systems configured with 8x H100 NVL outperform HGX A100 systems by up to 12X on GPT3-175B LLM throughput.
Note – this GPU must be installed in compatible servers. Below are popular H100 NVL servers:
8 GPU PCI-e H100 NVL Server (EPYC)