GPU and Accelerator Servers: Power, Cooling and PCIe Limits
Adding accelerators to a server is rarely a matter of fitting a card. The card is the easy part; the chassis, power, cooling and PCIe budget are what decide whether it is possible at all.
The four constraints
Before choosing an accelerator, establish whether the server can host one.
Physical space. Most accelerators are full-height and many are double-width, occupying two slot positions. A 1U chassis generally cannot take them at all, and 2U takes a limited number. Multi-accelerator systems are usually 4U for this reason.
Power. Accelerators draw substantial power, frequently more than everything else in the chassis combined. That has to come from the power supplies, and many cards need supplementary power connectors the chassis may not provide.
Cooling. All that power becomes heat. Server-grade accelerators are usually passively cooled, relying entirely on chassis airflow, which means the chassis must be designed to move air through them. Fitting a passively cooled card into a chassis not designed for it produces throttling or shutdown.
PCIe lanes. Each accelerator consumes lanes from a finite budget determined by the processors, and it competes with network cards, HBAs and NVMe drives.
All four have to work. A card that fits physically but exceeds the power budget is not usable.
Power in detail
This is where retrofits most often fail.
Adding accelerators to an existing server frequently requires higher-capacity power supplies, which are model-specific and must match each other — many platforms require redundant supplies to be identical wattage. Our power supply guide covers this.
Then the rack has to supply it. A server drawing considerably more than it did changes the circuit load, and racks reach power limits before space limits. In a redundant power design each circuit must carry the whole load alone, so the headroom is smaller than it appears. Our PDU guide covers the arithmetic.
Check the rack circuit capacity before ordering cards. Discovering the constraint afterwards means new circuits, which is slow and expensive.
Cooling in detail
Accelerator heat compounds every existing airflow weakness.
A rack that was borderline becomes a problem, and the failure is not dramatic — the accelerator throttles, delivering less performance than specified, and nobody notices because there is no error.
Three things to verify. That the chassis supports passively cooled cards in the positions you intend. That rack airflow is sound — blanking panels fitted, no recirculation, hot exhaust not feeding another intake. And that inlet temperature is within specification, measurable through the management controller.
Also expect noise. Chassis fans run faster to move the additional heat, and a server that was tolerable in a comms room may not be.
Memory and storage follow
Accelerators are rarely the only thing that needs scaling.
Data has to be fed to the accelerator, and that data comes from memory and storage. A system with substantial accelerator capacity and insufficient memory bandwidth or slow storage spends its time waiting.
Memory. Populate channels evenly for full bandwidth, and check the capacity ceiling for your register type. Our memory guide covers population rules.
Storage. Accelerator workloads frequently read large datasets repeatedly, which is a case for local NVMe. But NVMe consumes the same PCIe lanes the accelerators need, so the two compete directly. That trade-off has to be planned rather than discovered.
Processors. Lane count comes from the processor, so a processor with fewer lanes constrains what the system can hold regardless of physical slots. Our processor guide covers this.
Buy the platform or retrofit?
Worth being direct about, because retrofitting is frequently the wrong answer.
Retrofit works when the chassis already supports accelerators, the power supplies have headroom or can be upgraded, the rack circuit has capacity, and the PCIe budget allows it. Some servers are specified with this in mind.
Retrofit usually does not work on a 1U chassis, on a machine already near its power limit, where the chassis was not designed for passively cooled cards, or where lanes are already committed to storage and networking.
Where several of those apply, a platform designed for accelerators is cheaper than working around a chassis that was not — particularly once you count the upgraded power supplies, possible new circuits and the risk of throttling.
Practical checks before ordering
Confirm the chassis supports the card physically, in the specific slot positions, at full height and double width if applicable.
Confirm the power supplies can deliver the total draw, and whether higher-capacity units are needed and available for that platform.
Confirm supplementary power connectors are present if the card requires them.
Confirm the rack circuit has capacity, allowing for redundant paths each carrying the whole load.
Confirm the PCIe lane budget covers the cards plus existing network and storage devices.
Confirm cooling — passive card support, airflow soundness, inlet temperature.
Send us the server model and generation and what you intend to fit, and we will confirm whether the platform supports it before you order.
Common questions
Can I add an accelerator to my existing server?
Only if four things allow it: physical space for a full-height double-width card, power supply capacity, chassis cooling designed for passively cooled cards, and available PCIe lanes. A card that fits physically but exceeds the power budget is not usable.
Why can a 1U server not take accelerators?
Most accelerators are full-height and often double-width, which 1U cannot accommodate, and 1U also has the least cooling headroom. Multi-accelerator systems are typically 4U for both reasons.
What happens if cooling is inadequate?
The accelerator throttles, delivering less than specified performance with no error raised, so nobody notices. Server-grade cards are usually passively cooled and rely entirely on chassis airflow, so the chassis must be designed to move air through them.
Do accelerators compete with NVMe storage?
Directly. Both draw from the same finite PCIe lane budget determined by the processors, alongside network cards and HBAs. Accelerator workloads often want local NVMe, so the trade-off has to be planned rather than discovered.
Should I retrofit or buy a purpose-built platform?
Retrofit where the chassis already supports accelerators and power, cooling and lanes allow it. Where several constraints bind, a purpose-built platform is usually cheaper once you count upgraded power supplies, possible new circuits and the risk of throttling.
Send us the server model and what you intend to fit and we will confirm whether the platform supports it.
