PSUs, rails and risers: matching parts to a chassis
A server is only as good as the parts that feed it power, hold it in the rack, and connect its expansion cards. These are the three areas where a chassis that looks right on paper quietly refuses to accept an upgrade — or arrives missing the bits that make it usable. Here is how to match them correctly.
Power supplies: redundancy, sizing and efficiency
Redundancy. Enterprise servers take two hot-plug power supplies in a 1+1 redundant pair: either unit can fail, or be swapped, without downtime, and by default the load is shared across both. The rule that catches people out is that a redundant pair must be identical — the same wattage and the same efficiency rating. Mixing a 750 W and a 1100 W unit, or a Platinum with a Titanium, forfeits redundancy on most platforms. In a South African context, pairing the two PSUs across separate power feeds / UPS paths is what lets the server ride through a feed failure rather than just a PSU failure.
Sizing. The PSU has to cover the fully configured load — CPUs, every DIMM, all drives and especially any GPUs — with headroom to spare. A base config might be happy on a modest PSU; add high-wattage accelerators and you are pushed into the higher-wattage classes and, for full rated output and top efficiency, onto 200–240 V input (which suits SA's 230 V mains).
Efficiency (80 PLUS). The 80 PLUS programme certifies how much mains power a PSU delivers as usable DC rather than waste heat, measuring efficiency at 10%, 20%, 50% and 100% of rated load. Server power supplies are certified on the 230 V internal-redundant test. Higher tiers waste less energy and throw off less heat — both of which lower running cost and cooling load.
| 80 PLUS tier | Approx. peak efficiency (230 V internal-redundant test) |
|---|---|
| Gold | ~92% |
| Platinum | ~94% |
| Titanium | ~96% |
Titanium is the highest tier and holds the most demanding efficiency bar across the whole load range, so it stays efficient even on lightly-loaded servers (80 PLUS programme, CLEAResult; 80 PLUS efficiency levels). Dell offers hot-plug redundant PSUs up to Titanium (its Titanium units run on 200–240 V AC input) (Dell 750 W Titanium PSU, 200–240 V AC), and HPE states its Flexible Slot power supplies reach up to 96% efficiency (Titanium) with tool-less hot-plug and load-balanced redundancy (HPE Flexible Slot power supplies).
Rails: static vs sliding, and rack depth
Rails are not an afterthought — the wrong kit means the server will not mount, or cannot be serviced in place.
- Static rails fix the server in the rack. They are cheaper and lighter, but to service the machine you must slide it fully out and support its weight, and you cannot extend it while it stays cabled.
- Sliding rails (Dell markets these as ReadyRails) let the server slide out on rails for servicing without removing it, and they support a Cable Management Arm (CMA) so a fully-cabled server can be pulled into a service position without unplugging anything. They mount tool-lessly to square- and round-hole racks, with adapter brackets for threaded-hole racks.
The critical compatibility factor is rack mounting depth. Rails adjust over a range, and a rack that is too shallow or too deep will not accept a given kit — Dell publishes a Rail Sizing and Rack Compatibility Matrix for exactly this, grouping each kit into a family identified by a code (Dell Rail Sizing and Rack Compatibility Matrix). Second-hand servers very often arrive with the wrong rails or none at all, so confirm the kit matches both the server model and your rack's depth and hole type, and choose sliding + CMA if you will service the machine in place.
Risers and GPU enablement: the empty slot is not enough
A server's PCIe expansion slots are presented through risers — cards that route the mainboard's PCIe lanes into usable slots. Different riser configurations change how many slots you get, their form factor (full-height/half-height, full-length/half-length) and their lane width (x8/x16). Which riser is fitted therefore decides what cards will physically and electrically fit.
GPUs and other accelerators need more than a slot. They typically require:
- a riser configuration that can physically hold a double-width card and give it x16 lanes,
- auxiliary GPU power cabling (6+2-pin, 8-pin or 16-pin),
- GPU air shrouds and high-performance fans to move enough air, and
- PSU headroom to feed the card.
Vendors bundle these as a GPU enablement kit. Dell's PowerEdge R750 GPU kit, for instance, adds specific full-length and half-length risers, a GPU air shroud with high-performance heat sinks, six high-performance fans, and the 6+2-pin / 8-pin / 16-pin power cables — supporting a maximum of two double-width or eight single-width GPUs, where low-power cards (e.g. NVIDIA T4/A2) need no auxiliary power while higher-wattage cards (e.g. H100/L40) require the 16-pin auxiliary power cabling (Dell PowerEdge R750 GPU kit).
Why it matters: a chassis that "supports a GPU" on the datasheet still needs the riser, shroud, fans, power cables and PSU capacity actually fitted. When buying refurbished for accelerated workloads, confirm the full enablement kit is present — not just an empty slot and an optimistic spec sheet.
Server Hub builds and checks these details in-house — matched PSUs, the right rail family for your rack, and complete GPU enablement — on the servers in our catalogue.
The Server Hub briefing
South African IT hardware news, once a week.
What’s new, what it costs in rand, and what it means for the kit you run — servers and storage, networking, backup power, surveillance and print. Every claim checked against a named source.