NVIDIA
Model H100
The reference training accelerator for multi-GPU nodes, deployed eight to a baseboard with full NVLink bandwidth between every GPU.
Quantity
No prices online. Add what you need and we reply with an all-in price and a delivery date.
| Architecture | Hopper |
|---|---|
| GPU memory | 80GB HBM3 |
| Memory bandwidth | 3.35TB/s |
| Interconnect | NVLink, 900GB/s |
| Form factor | SXM module on HGX baseboard |
| Max power | Up to 700W |
Specifications are indicative and confirmed on quote. Product and brand names belong to their respective owners and are used for identification only.
The questions we ask before quoting. Answer them in the line notes and the first reply is usually the final one.
NVIDIA
AI AcceleratorsHopper-generation accelerator with the largest memory pool in its family, suited to serving and fine-tuning large language models that outgrow 80GB parts.
NVIDIA
AI AcceleratorsA PCIe card version of Hopper with extra memory, built for LLM inference in standard air-cooled servers and bridged in pairs over NVLink.
NVIDIA
AI AcceleratorsA general-purpose data-centre GPU that covers inference, fine-tuning, rendering and virtual workstations from one PCIe card.
