Adding a GPU to a Server: Risers, Power, Cooling and Lanes
Fitting a GPU or accelerator into a server is the upgrade most likely to be stopped by something other than the card. This covers the four constraints to check before ordering.
The four things that decide whether a GPU fits
In the order they stop purchases.
Physical space. Full-height, full-length, double-width cards need a chassis and riser arrangement that accommodates them. A 1U server takes low-profile cards only, and most accelerators are not low-profile. Our form factors guide covers what each height allows.
Power. Two separate questions: whether the power supplies can carry the additional draw, and whether the chassis provides the supplementary power connectors the card needs. Many server chassis have specific GPU power cables that are a separate part.
Cooling. Datacentre accelerators are frequently passively cooled, relying on chassis airflow rather than their own fans. That works only in a chassis designed for it, with the correct air baffle and frequently higher-performance fans. Our fans guide covers this.
PCIe lanes. A card needing sixteen lanes in a slot wired for eight runs at half bandwidth, with nothing reporting a fault. Our PCIe guide covers reading the actual link width.
The riser question
Most rack servers put expansion cards on risers rather than directly on the board, and the riser determines what fits.
Three things follow. Risers are chassis and generation specific. A server may support a GPU only with a particular riser that was not fitted. And GPU-capable risers are frequently different parts from standard ones, because they route more lanes and provide the mounting for a longer card.
So the question to answer before ordering is not "does this server support this GPU" but "does this server, with the riser it has, support this GPU."
Passive versus active cooling
The distinction that causes most failed installations.
Passively cooled accelerators have no fan. They are designed for server chassis with high static-pressure fans pushing air through them. In a workstation or a chassis without that airflow, they overheat and throttle or shut down.
Actively cooled cards have their own fans and work in workstations and towers, but exhaust into the chassis rather than out of it, which matters in a dense rack.
The practical rule: a passive datacentre card belongs only in a chassis rated for it. Our T4 accelerator guide covers one example, and the T4 is unusual in being low-power enough to suit more chassis than most.
Power supply sizing
Two checks.
Total draw with the card fitted, against the supply rating. A GPU can add several hundred watts, and a chassis sized for its original configuration may not have the headroom. Our power supply guide covers matching a larger supply.
Redundancy still works. If the machine has two supplies, each must carry the full load alone, including the card. A GPU can turn a genuinely redundant machine into one that only appears redundant.
And beyond the machine, the rack circuit and the cooling. Several GPU-equipped servers in a rack can exceed what the circuit or the room was planned for. Our server room guide covers the heat side.
What to confirm before ordering
Six things, and a vendor configuration guide answers most of them.
The chassis supports the card in the vendor’s own documentation, not by inference.
The riser needed, and whether you have it.
The GPU power cable, which is a separate chassis-specific part.
The air baffle or shroud variant required.
Higher-performance fans, if the platform requires them for GPU configurations.
Power supply capacity with redundancy intact.
Notice that five of the six are not the GPU. This is a components exercise, and our build guide covers the same pattern across other upgrades.
Used accelerators: what to check
Four things.
What it was used for. Cards run continuously at full load for years have had a harder life than the hours alone suggest.
Thermal condition. Dust in the heatsink and degraded thermal interface material both raise temperatures, and both are fixable if you know to check.
That it is the variant you expect. Accelerators ship in passive and active versions, and in different memory configurations, under similar names.
Driver and software support status for the generation, which on accelerators moves faster than on general server hardware.
Frequently asked questions
Will a GPU fit in my server?
Four things decide it: physical space and riser arrangement, whether the supplies can carry the draw and provide the right connectors, whether the chassis can cool it, and whether the slot has the PCIe lanes the card needs.
Why does my accelerator overheat?
Datacentre accelerators are frequently passively cooled, with no fan, relying on chassis airflow. In a workstation or a chassis without high static-pressure fans and the correct baffle, they overheat and throttle. A passive card belongs only in a chassis rated for it.
Do I need a special riser for a GPU?
Frequently yes. Risers are chassis and generation specific, and GPU-capable ones are often different parts because they route more lanes and mount a longer card. The question is whether the server with the riser it has supports the card.
Will a GPU break my power redundancy?
It can. Each supply must carry the full load alone, including the card. A GPU adding several hundred watts can turn a genuinely redundant machine into one that only appears redundant.
Why is my GPU running at half speed?
Probably PCIe lanes. A card needing sixteen lanes in a slot wired for eight runs at half bandwidth with nothing reporting a fault. Check the actual negotiated link width rather than the physical slot size.
What should I check on a used accelerator?
What it was used for, its thermal condition including dust and thermal interface material, that it is the variant you expect since passive and active versions ship under similar names, and driver support status for the generation.
Tell us the server model and the card, and we will confirm the riser, power cable, baffle and fans it needs before you order any of them.
