Buying in volume? Request a bulk quote
We source this item through our supplier network and do not hold inventory. Price and availability are confirmed with you before any order is accepted.
Questions about this part? Talk to us.
Phone +1 872 293 4196 · Email supply@allchipsupply.com
Location: USA · Shipping: Worldwide Shipping
The NVIDIA H100 NVL is a PCIe version of the Hopper data center GPU built for large language model inference. Each card carries 94 GB of HBM3 memory with about 3.9 TB/s of bandwidth, according to NVIDIA's specification, and connects over PCIe 5.0 x16.
Two cards can be paired with an NVLink bridge, giving a two-card set 188 GB of GPU memory. Maximum power is configurable from 350 to 400 W, and Multi-Instance GPU can create up to seven instances per card. The card is dual-slot, full-height and full-length, and it is passively cooled, so it needs a server with suitable front-to-back airflow. Confirm the server is qualified for H100 NVL.
| Architecture | NVIDIA Hopper |
|---|---|
| GPU memory | 94 GB HBM3 |
| Memory bandwidth | 3.9 TB/s (NVIDIA spec) |
| Interface | PCIe 5.0 x16 |
| Form factor | Dual-slot, full-height full-length, passive |
| Max power | 350-400 W (configurable) |
| Multi-GPU | NVLink bridge (2-card) |
| Multi-Instance GPU | Up to 7 MIG instances |
Specifications are summarized from manufacturer and listing information; confirm against the manufacturer datasheet before purchase.
Manufacturer datasheet / product page (opens the manufacturer's website)
NVIDIA Hopper · 80 GB HBM2e · 2.0 TB/s
NVIDIA Hopper · 141 GB HBM3e · 4.8 TB/s
200 Gb/s · PCIe 5.0 x16 · 1
NVIDIA Ampere · 80 GB HBM2e, ECC · ~1.9 TB/s
