Add a GPU to an existing rack server: L4 or RTX PRO 4500 Server Edition, the checklist

Which GPU fits a server you already run

To add a GPU to a rack server that already runs virtual machines or applications, start with the NVIDIA L4: at 72 W it draws its power from the PCIe slot, needs no power cable, is passive, low profile and single slot, and the large server makers list it for many of their general-purpose models. The RTX PRO 4500 Blackwell Server Edition brings 32 GB and Blackwell features at 165 W, but needs a full-height, full-length slot position and a 16-pin power cable made for that server. Whichever card you choose, the exact server model, riser set and slot have to be on the server maker's support list.

This page covers the general-purpose host: a 1U or 2U server bought for virtualisation, databases or file services that should now accelerate inference, video or virtual desktops. Large cards have their own checklists: the H200 NVL in an existing server and the RTX PRO 6000 Server Edition in our list of compatible servers. Which card replaces an older T4 is the subject of our T4 successor comparison.

NVIDIA product pages and vGPU list as summarised in our research notes of 24 October 2026; Lenovo Press LP1717 (L4, 15 September 2024) and LP2391 (RTX PRO 4500 Server Edition, 28 July 2026), read on 25 October 2026.

The RTX PRO 4000 SFF is in the table because it also runs from slot power at 70 W, but it is a workstation card with its own fan and two slot widths, and we found no server maker that lists it for a rack server. For a rack host, plan with the L4 or the RTX PRO 4500 Server Edition. Our comparison of the L4 and RTX PRO 4500 Server Edition weighs the two for language models, video and vGPU.

Check 1: the maker's support list and riser configuration for your model

A server maker supports a GPU in a configuration: the model, the riser set, the slot, the cable and sometimes the fan set. The empty slot at the back of the chassis says little about it. Dell's technical guide for the PowerEdge R770, Rev. A09 of June 2026, shows how much the riser decides.

Dell PowerEdge R770 Technical Guide, part number E111S, Rev. A09, June 2026, Table 29 "NVIDIA GPU support list", read on 25 October 2026. All entries single width, PCIe x16. Dell lists the card as "RTX Pro 4500" without the edition; its 165 W rating is that of the Server Edition.

A server bought with riser configuration 1 therefore takes two L4, and the same model with configuration 12 takes six. If the riser set has to change, the riser kit, its cables and the slot mapping follow the maker's parts list. Dell's guide also states: "Do not install GPUs, network cards, or other PCIe devices on your system that are not validated and tested by Dell."

Lenovo publishes the counts per server in the product guide of each GPU. For the L4, guide LP1717 lists, among current models, up to 3 in the ThinkSystem SR630 V3 and SR630 V4, 8 in the SR650 V3 and SR650a V4, 10 in the SR650 V4, 4 in the SR635 V3 and SR645 V3, 8 in the SR655 V3 and 7 in the SR665 V3. The single-socket ST50 V3, ST250 V3 and SR250 V3 are marked as not supported. For the RTX PRO 4500 Server Edition, guide LP2391 of 28 July 2026 gives power cables for the SR650a V4 and the SR665 V3, and Lenovo's ServerProven site is the reference for which other models and counts are supported.

Look up your server by its full machine type, not by the family name: Lenovo lists the H200 NVL for the SR650a V4 and not for the SR650 V4, as our H200 NVL checklist shows. HPE publishes its GPU lists in its NVIDIA Accelerators QuickSpecs and the QuickSpecs of each ProLiant model.

Check 2: slot power and the auxiliary cable

A PCIe slot delivers up to 75 W; Dell's R770 guide speaks of "6 x single width 75 W accelerators". The L4 at 72 W fits that budget, and Lenovo's guide states: "The NVIDIA L4 does not require an auxiliary power cable." That is why the L4 is the card for hosts without GPU power cabling.

The RTX PRO 4500 Server Edition draws 165 W and takes the rest through a 16-pin connector. Lenovo's guide states that the option "does not ship with auxiliary power cables", that cables are server-specific, and that a field upgrade needs the cable ordered separately. It names, for example, a 320 mm 16-pin (2x6+4) cable kit, 4X97B04860, for the SR650a V4, and a 400 mm cable for the SR665 V3, three of which serve one to three GPUs; with four GPUs that server uses two single cables and one 400 mm Y-cable for two GPUs, and Lenovo lists a separate thermal kit, 4X67A85856, for the card in that server. For the SR630 V4 and SR650 V4 the guide lists no cable, so ask Lenovo before you order. A server without the matching cable or connector on its board cannot take the card.

Then add up the power supplies. By our arithmetic, six L4 draw 432 W and four RTX PRO 4500 Server Edition 660 W of board power, on top of what the server draws today. With redundant supplies, one supply must carry the whole server. The server maker's power calculator gives the figure for your configuration; our guide to GPU server power on 16 A, 32 A and three-phase feeds covers the rack side.

We supply the L4 and the RTX PRO 4500 Server Edition for servers you already run and check the platform, power and cooling first; send us the exact server model, its riser configuration and the free slots, and our team replies with the card and count that fit.

Check 3: slot height, length and bracket

The L4 is a low-profile card of 169 × 69 mm. Lenovo's option ships with a full-height bracket attached and a low-profile bracket in the box, so the same card fits both slot heights. The RTX PRO 4500 Server Edition is 4.4 inches high and 10.5 inches long, "Single Slot, Full Height, Full Length" in Lenovo's words, and needs a full-height riser position with that length free, which a low-profile slot does not offer. Dell's table lists every GPU in an x16 slot.

Check 4: airflow and temperature

Both server cards are passive. Lenovo lists the L4's thermal solution as "Passive" and describes the RTX PRO 4500 Server Edition as "passively cooled", needing the system's airflow to operate. Lenovo gives both cards an operational ambient of 0 °C to 50 °C, but that is the card's own rating; the server's thermal rules for the configuration are stricter and decide.

Some platforms require extra parts for GPUs: Lenovo lists a thermal kit for the RTX PRO 4500 Server Edition in the SR665 V3, and Dell's R770 guide requires High Performance Platinum fans for the H200 NVL and the RTX Pro 6000 it lists at 450 W. For 72 W and 165 W cards, read the server's thermal section for fan type, air ducts and drive configuration before you order. Our guide to GPU server inlet temperature explains how to measure the inlet at the rack position.

Check 5: firmware, operating system and hypervisor

Bring the server's BIOS and management controller firmware to the level the maker lists for the GPU before the card goes in. Then install the NVIDIA driver from the branch your software needs; our overview of NVIDIA driver branches and CUDA versions lists them, including the open kernel modules NVIDIA requires for Blackwell cards such as the RTX PRO 4500 Server Edition. With Secure Boot on, the kernel modules have to be signed, as our guide to the NVIDIA driver with Secure Boot shows.

Many general-purpose hosts run a hypervisor, and there are two ways to give the card to virtual machines.

Our guides to GPUs in vSphere and vGPU on Proxmox VE, from NVIDIA's vGPU documentation (vGPU 20.2, August 2026) and Broadcom's DirectPath I/O documentation, read on 24 October 2026.

On vSphere, NVIDIA lists the L4 for vGPU from release 15.2 and the RTX PRO 4500 Server Edition from 20.0, and Blackwell Server Editions need ESXi 8.0 Update 3g (8.0 U3 P06) or later, so an older host is updated first. NVIDIA vGPU needs the vGPU Manager on every GPU host, and a vGPU machine moves by vMotion only to a host with the same GPU type. A passthrough VM loses vMotion, snapshots while running and HA, as our guide to GPUs in vSphere: passthrough or vGPU explains. On Proxmox VE, NVIDIA's R595 documentation for Linux with KVM lists both cards; our guide to NVIDIA vGPU on Proxmox VE covers the setup. Profiles per card are in our L4 vGPU profiles and RTX PRO 4500 Server Edition vGPU profiles guides.

Check 6: warranty and support

The makers state their position for their own option part numbers. Lenovo's guides for both cards give a "One year limited warranty" and add: "When installed in a Lenovo server, the GPU assumes the server's base warranty and any warranty upgrades." Dell's caution against devices "not validated and tested by Dell" points the same way: keep the card, riser, cable and fan set within the maker's list.

For a card bought separately from the server maker, ask the maker before you order how it handles support calls on that server, and keep the order documents. The cards we supply carry the manufacturer's warranty, handled through us.

When a new GPU server is the better path

A general-purpose host is a good place for one to six small cards. A new server is the better choice when the host's model or riser set has no GPU listing, when the workload needs more than 32 GB per card or 600 W cards, or when the GPUs would compete with the host's existing virtual machines for power and cooling. Our pages on buying an L4 server and the eight-card L4 server describe purpose-built options with the same card, and our guide to how many GPUs fit in one server explains lanes, power and airflow per chassis.

What we supply

We supply the NVIDIA L4 and the RTX PRO 4500 Server Edition, and the other professional NVIDIA cards on our GPU page, as cards for servers you already run, on one EU contract and invoice with manufacturer warranty, and build AI servers to order when a new server fits better. Server option parts such as risers, cables and fan kits come from the server maker. Quantities and lead time are confirmed in writing with the quote.

FAQ