What Is a GPU Server? Accelerated Rack Servers Explained

An ordinary 1U or 2U server is built to hold drives and network cards. Its expansion slots are usually single-width, its risers carry modest lane counts, and its power budget assumes nothing in a slot draws more than a few dozen watts. That is a perfectly good host for a network adapter or a storage controller and a poor one for an accelerator.

A GPU host inverts those assumptions. It provides full-height, full-length double-width slots, routes a full x16 link to each of them, and sizes the power supplies and the fan wall for cards that can each draw several hundred watts. NVIDIA quotes a PCIe Gen4 x16 link at 64 GB/s, and states that PCI Express Gen 5 provides double the bandwidth of Gen 4 — which is why the PCIe generation of the host is part of the accelerator's own specification, not a footnote.

Slot width comes first because it is physical. NVIDIA's L4 is a single-slot, low-profile card, so it fits hosts that were never intended for accelerators and can be installed several to a chassis. The RTX PRO 6000 Blackwell Server Edition is, in its air-cooled form, a dual-slot full-height full-length card; the liquid-cooled variant is single-slot but full-height extra-long, which is a different chassis question rather than an easier one.

Power comes second and is the constraint that most often decides the count. The L4 is rated at 72 W maximum. The RTX PRO 6000 Blackwell Server Edition is configurable up to 600 W. Eight of the first is a different power design from two of the second, and the rack circuit behind the host has to carry whichever you choose.

Lanes come third. Each accelerator wants its own x16 link, and a two-socket host has a finite number of lanes to distribute across accelerators, network adapters and storage. A configuration that looks fine on a slot count can still be bottlenecked by how the risers are wired, which is a question to put to the host vendor rather than to infer from a product photo.

Data-center accelerators are sold in air-cooled and liquid-cooled forms, and the difference is not only thermal. NVIDIA lists the RTX PRO 6000 Blackwell Server Edition as dual-slot full-height full-length for air and single-slot full-height extra-long for liquid — so the two variants of one card occupy different physical envelopes and belong to different chassis.

Liquid cooling raises the density a rack can carry, and it introduces a facility dependency: manifolds, coolant distribution and a service procedure that the room has to support. For a first accelerated host, the air-cooled variant in a qualified chassis is usually the shorter path, and the decision is worth making before the order rather than after the delivery.

Start with the host vendor's qualified list for the exact card, in the exact variant, at the quantity you want. Confirm the slot form factor, the PCIe generation and lane width per slot, the per-slot power delivery, and whether the chassis needs a higher-airflow fan option or a different heatsink kit when accelerators are fitted.

Then check the software. NVIDIA AI Enterprise is licensed per GPU — a license for every GPU in the server that hosts the software — so the accelerator count sets the license count. Finally, size the rack: total the host's power draw with the cards fitted, compare it against the circuit and the PDU feeding the cabinet, and confirm the depth and weight the chassis adds before it arrives.