Adding GPUs to a server later: what a scalable GPU server needs from the first day

Adding GPUs to a server later: what to plan on day one

You can add GPUs to a server later, from two cards to four or eight without a second server, if the server was ordered as the larger configuration from the start. The base model, the number and generation of processors, the system memory, the power supplies, the fan set and the riser and cable kits all have to match the final card count. The server maker also has to list that count for the exact model and configuration.

Server makers treat a later GPU upgrade as a change of configuration. Dell’s technical guide for the PowerEdge XE7745, Rev. A09 of September 2026, says: “Do not install or remove GPUs without first consulting Dell.” The same guide states that damage from hardware installation that is not authorised and validated voids the system warranty. Lenovo’s SR650a V4 guide requires a DIMM filler in every empty memory slot when a GPU is added as a field upgrade.

Read on 10 October 2026: Lenovo Press LP1611 (updated 8 October 2026), LP2128 (5 October 2026), LP2263 (28 July 2026) and LP1944 (7 November 2025); NVIDIA’s configuration guide for NVIDIA-Certified Systems (30 September 2026) and AI Enterprise licensing guide (2 September 2026); Dell XE7745 Technical Guide Rev. A09 and R770 Technical Guide Rev. A09.

Chassis and base model: the maximum card count is fixed at order

Lenovo’s product guide for the ThinkSystem SR675 V3, updated on 8 October 2026, lists the server in three base configurations. Besides an SXM5 model there is a “4-DW GPU model (4x double-wide or single-wide PCIe GPUs with 8x 2.5-inch drive bays)” and an “8-DW GPU model (8x double-wide or single-wide PCIe GPUs with EDSFF drive bays)”. The guide says nothing about converting one into the other. A company that expects eight cards orders the 8-DW model, even with two cards at first.

The 2U SR650a V4 carries “up to 4x double-wide GPUs or 8x single-wide GPUs, installed in the front slots”. Its storage configuration table lists four double-wide GPUs with one or with two processors, depending on the drive configuration, and notes that “Configurations listed with 2xDW GPUs are for configurations with 600W GPUs.” Lenovo lists the RTX PRO 6000 Server Edition for this server as a 600 W card and as a card “Slot Capped at 450W”, and its change history of 28 July 2026 adds the capped version “to allow 4x GPUs to be installed in the SR650a V4”. A server meant to reach four cards is therefore ordered with the capped version and a drive configuration listed for four double-wide GPUs.

Dell’s PowerEdge Server GPU Matrix of August 2026 notes that the “maximum number of GPUs allowed might differ in different configurations on the same platform”. Our article on how many GPUs fit in one server explains the lane, power and airflow limits behind these counts, and the list of servers compatible with the RTX PRO 6000 Server Edition gives the maximum per model.

Processors and system memory for the final GPU count

NVIDIA’s configuration guide for NVIDIA-Certified Systems, last updated on 30 September 2026, calls “2x / 4x / 8x GPUs per server” a balanced configuration. It adds that “GPUs should be evenly distributed across CPU sockets and PCIe root ports”, and it asks for a PCIe Gen5 x16 link for each RTX PRO 6000 or H200 NVL and a minimum of two CPU sockets. Its minimums grow with the card count: six physical CPU cores per GPU and system memory of at least twice the total GPU memory. NVIDIA presents these figures as “a starting point for addressing workload-specific needs”.

By our arithmetic, two RTX PRO 6000 cards hold 192 GB of GPU memory, which gives a minimum of 384 GB of system memory and 12 cores. Eight cards hold 768 GB and give 1,536 GB and 48 cores.

Lenovo’s SR650a V4 guide states “All installed DIMMs must be identical part numbers; mixing not supported”, so modules added later must match the installed ones. Order the memory for the final count, or larger modules with slots left empty.

Lenovo supports the RTX PRO 6000 in the SR675 V3, up to eight cards, “with 5th Gen AMD EPYC processors only”, so an SR675 V3 ordered with an earlier EPYC generation is outside that support.

Power supplies sized for the final GPU count

Dell’s XE7745 splits its power supplies into a CPU zone and a GPU zone, with six of its eight bays in the GPU zone. Its guide says: “To ensure full redundancy, install N+N number of PSUs in each zone, i.e., 1+1 in CPU zone and 3+3 in GPU zone.” With four supplies, Dell lists 600 W cards only in the non-redundant layout, and eight of them only on 3,200 W supplies. With eight 3,200 W supplies, all three layouts are supported for two, four and eight cards, A/B grid included. With eight 2,400 W supplies, eight 600 W cards are supported without redundancy and in the 7+1 layout, but not in the A/B grid layout.

Upgrading the wattage later means replacing supplies. Dell’s guide states that “the wattage capacity of the PSU currently in use must match the newly added PSU to enable it”. A server bought with two cards and 2,400 W supplies on A and B feeds therefore cannot keep A/B grid redundancy at eight cards unless all its supplies are exchanged. The SR650a V4 holds at most two power supplies, with options from 800 to 3,200 W, so the wattage ordered for two cards stays the wattage for four unless both supplies are replaced.

Four 600 W cards draw 2.4 kW and eight draw 4.8 kW before processors and fans, so plan the rack feed for the final count too.

We build AI servers to order and check the rack, power and airflow before we quote. Tell us how many cards you start with and how many you expect to add, and we size the CPU, memory, power and cooling for that count.

Fans, risers, power cables and fillers

Passive GPUs take their air from the server’s fans. Dell’s R770 Technical Guide, Rev. A09 of June 2026, lists minimum requirements for the H200 NVL and the RTX PRO 6000: “All fan modules must be High performance Platinum Fan (HPR Platinum) type.” The same guide states that “The puck power cables that are required for GPU installations are specific to the slot”.

For a server that will grow, order the high-performance fan set and the GPU riser cages for the final card count with the first configuration. Record the power cable part number for each empty GPU slot, because the cable is ordered per server and slot. Our checklist for an H200 NVL in an existing server covers the cable sense pins, the BIOS settings for large memory windows and the server makers’ support lists for that card.

NVLink bridges and MIG: decide the topology before the cards arrive

NVIDIA’s H200 page gives the H200 NVL a “2- or 4-way NVIDIA NVLink bridge: 900GB/s per GPU”, and Lenovo sells the two bridges as separate parts, 4X67A97320 for 2-way and 4X67A97322 for 4-way. NVIDIA’s developer blog of 11 December 2024 gives four bridged cards a “combined 564 GB of HBM3e memory”. A 4-way bridge spans four cards, so a plan to run one model across four H200 NVL needs a server whose maker supports the 4-way bridge in the slots you will use. By our reading, two cards bought first on a 2-way bridge and two added later form two 2-way pairs unless the bridge is replaced. Models that fit on one card and run as copies need no bridge.

The RTX PRO 6000 Server Edition has no NVLink, so adding cards does not change any bridge plan. NVIDIA’s page for the card describes MIG as “enabling the creation of up to four fully isolated instances”, and the H200 page gives the H200 NVL “Up to 7 MIGs @16.5GB each”. MIG divides one card, so each added card brings its own set of instances.

The Dell XE7745 and R770 guides, Dell’s GPU matrix and Lenovo’s SR650a V4 and SR675 V3 guides state no rule on mixing GPU models in one server; they list maximum quantities per card model. We recommend growing with the model the server started with, because the NVLink bridges, the MIG profiles and the power, cooling and cable rules are all defined per card model.

NVIDIA AI Enterprise licences as cards are added

NVIDIA’s licensing guide, updated on 2 September 2026, states that “NVIDIA AI Enterprise software is licensed on a per-GPU basis.” Each added card therefore adds a licence wherever the server runs AI Enterprise software, such as NIM in production or compute vGPU profiles. Our article on how NVIDIA AI Enterprise is licensed explains which deployments need it.

The H200 NVL is licensed differently. NVIDIA states that “Each NVIDIA H200 NVL Tensor Core GPU includes a five-year NVIDIA AI Enterprise subscription”, that its start date is based on the ship date of the GPU board to the server maker plus 90 days, and that it “cannot be modified, as it is tied to the specific card”. Two cards whose boards shipped to the server maker a year after the first two therefore end their subscriptions about a year later. For the RTX PRO 6000, L40S and L4, which include no subscription, each added card adds one licence where the software runs.

Growth paths from two to eight GPUs

Each example starts from two cards and lists what the cited documents require on day one.

Lenovo Press LP2128, LP1611 and LP2263; Dell XE7745 Technical Guide Rev. A09, Tables 28 and 29; NVIDIA H200 page and Lenovo LP1944 for the bridges, read on 10 October 2026. Each server maker’s configurator decides per configuration.

We supply cards for servers you already run and check the platform, power and cooling first. Describe your server and the cards you plan to add in the form below.

Growing beyond one server

A second server is the next step once the first reaches its listed maximum, eight cards in the XE7745 and the SR675 V3, or when the service has to keep running while one server is down. It brings its own processors, memory, power supplies and rack position. A model that fits in one server runs as a full copy on each, behind a load balancer, and our article on LLM high availability with two GPU nodes sizes that layout. NVLink bridges join cards inside one server, so a model split across two servers communicates over the network. NVIDIA’s enterprise reference architecture for H200 NVL clusters uses a “2-8-5” configuration per server, where the digits refer to “the number of sockets (CPUs), the number of GPUs” and the network adapters.

What we supply

We build AI servers to order with the RTX PRO 6000 Server Edition, the H200 NVL with NVLink bridges, the L40S or the L4, assembled and burn-in tested, with manufacturer warranty on every component and on one EU contract and invoice. We check the rack, power and airflow before we quote. For a server you already own, we check the platform, power and cooling before you add cards and say in advance if something will not work. NVIDIA AI Enterprise and vGPU licences come on the same invoice as the cards, and the professional NVIDIA GPUs we supply include the RTX PRO range from the 2000 to the 6000.

FAQ