GPUs & AI accelerator cards

7 products

Products

A graphics card or an AI accelerator adds compute to a machine you already run. The shelf covers three kinds of buyer. Data center teams get PCIe cards such as the H200 NVL with 141 GB of HBM3, the H100 80 GB and the Atlas 300I Duo for video inference. Studios and AI labs get workstation and gaming graphics cards, from the 96 GB RTX PRO 6000 to the RTX 5090. Edge builds get a 160 TOPS M.2 accelerator that fits a mini PC.

Frequently asked questions

What is the difference between an AI accelerator and a graphics card?

A graphics card has display outputs and drives screens as well as compute; the RTX 5090 and the RTX PRO cards are graphics cards. An accelerator such as the Atlas 300I Duo is built only for compute, here 280 TOPS of INT8 inference with on-card video codecs, and sits in a server slot without a monitor attached.

Which GPU for AI should I choose?

Memory comes first. The model and its batch must fit on the card. The H200 NVL carries 141 GB, the RTX PRO 6000 Max-Q 96 GB and the RTX PRO 5500 84 GB of ECC memory, the H100 80 GB and the RTX 5090 32 GB. Then check that the chassis has the right slot (PCIe 5.0 or Gen4) and enough power and cooling for the card.

Can an M.2 AI accelerator replace a GPU?

For inference at the edge, often yes. The M.2 card adds 160 TOPS to a mini PC or embedded controller through an M.2 slot, enough for vision analytics or a small local language model. Training and large models still need a full GPU.

Can I order a single card?

It depends on the model. The RTX 5090, the Atlas 300I Duo and the M.2 accelerator start from one unit, the H200 NVL from two, the RTX PRO 5500 from five, and the RTX PRO 6000 Max-Q and the H100 from ten. The minimum is shown on each product page.