Learn
Four things account for most of the confusion: which physical size fits, which generation of the specification a unit follows, which cable goes where, and what is inside that makes one better than another at the same wattage.
Nothing here assumes prior knowledge. Any term with a dotted underline explains itself where it stands, and the diagrams are protected against casual copying — worth saying plainly that this stops accidental reuse, not a screenshot.
Part one
Get this wrong and nothing else matters, because the unit will not go in the case.
Whatever your case takes — check the case specification first, not the power supply. ATX (150 × 86 × 140 mm) is the desktop default: cheapest per watt, widest choice, and the only practical option above about 1000 W. SFX and SFX-L exist for small cases and cost more per watt for less output.
An SFX unit will physically mount in an ATX bay with a bracket, but there is rarely a reason to: you would be paying more for a louder, smaller unit in a case with room for a better one.
SFX-L, essentially always. Same mounting face, 30 mm more depth, and that depth is what fits a 120 mm fan instead of a 92 mm one. A larger fan moves the same air more slowly, which is the whole of the noise difference between two otherwise identical units.
Because the cables have to go somewhere once the body is in. A case that accepts ATX will not necessarily accept a 180 mm-deep ATX unit, and the shortfall shows up as cables crushed against a drive cage rather than as a unit that will not fit.
Part two
One revision of the specification matters to a buyer today. The rest is history.
| Revision | What it added |
|---|---|
| ATX 2.0–2.2 (2003) | Two +12V rails, the 24-pin main connector |
| ATX 2.3–2.4 (2008) | Efficiency targets; a single strong +12V rail becomes normal |
| ATX 2.5x (2018) | Alternative sleep mode; the long plateau |
| ATX 3.0 (2022) | Defined behaviour under transient excursions, plus 12VHPWR |
| ATX 3.1 (2024) | 12V-2×6 connector; stricter rules about plug seating |
If you are buying a current graphics card, yes. If the machine has integrated graphics or an older mid-range card, it makes little difference and you should spend the money on build quality instead.
The gap between 2008 and 2022 in the table is not an oversight. Nothing much needed to change until graphics cards began presenting very short, very large excursions — which is the entire content of .
A modern graphics card can briefly draw close to twice its rated power. That is a , and it lasts under a millisecond — far too brief for a wattmeter to show, and long enough to trip a protection circuit.
So capacity has to cover the spike, not the average. An unit is specified to ride those out; an ATX 2.x unit of the same wattage is not, and may simply shut the machine down.
No. It is the wrong choice specifically when a high-draw graphics card is involved. For an office machine, a media box or a build on integrated graphics, a good ATX 2.x unit is better than a mediocre one — the generation is one criterion among many, not a gate.
Part three
Seven cables in the box, several that look alike, and two that will fit sockets they must not go into.
The 24-pin goes to the long socket on the right edge of the motherboard. The 8-pin CPU plug goes to the top edge beside the processor. Graphics cards take either 6+2-pin PCIe plugs or a single on a current card. Drives take SATA power; fans and pumps take Molex or SATA depending on the accessory.
No, and this is the mistake that damages hardware. Both are 8-pin, but the notch shapes differ and the pinouts are not the same. Forcing a PCIe plug into a CPU socket, or the reverse, can destroy the board.
Cables and sockets are labelled — CPU or PCIE. Read the label rather than the shape.
Never, even between two units from the same manufacturer. The connector on the power supply end is not standardised: two units can use an identical-looking socket with a completely different pinout, and the result is a short across the rails.
Prefer native cables to adapters on anything drawing real power, too. The reported melting incidents concentrated on adapters and partly seated plugs, not on the connector itself — push until you feel the latch.
Part four
Two units of identical wattage and identical certification can differ at every stage of the chain. That difference is what the rating on this site scores.
Three stages account for most of it. at stage two is standard now; passive PFC on a unit sold today is grounds for an automatic F here. switching at stage three is what makes Gold and better efficiency practical, and an older topology on a recent unit signals an old platform in a new box.
Stage five is the one that matters most in a modern machine: conversion holds +5V and +3.3V independently, and cannot. Since a modern PC draws almost everything from +12V, is a real defect rather than a historical curiosity.
The bulk capacitor is the largest cylinder on the board and its maker is printed on the sleeve. It is the one component identifiable from a teardown photograph without any electronics knowledge, which is why reviewers always mention it — and why it is a reasonable proxy for how seriously the rest of the build was taken.
What you want is a 105 °C rating from a Japanese or reputable Taiwanese maker. An 85 °C part in that position is a corner cut in the hottest place in the unit.
Japanese — the benchmark
What a premium unit uses throughout. Seeing one on the bulk cap and 85 °C parts on the secondary side is a common cost-saving trick, so check both.
Taiwanese — good, in the right series
Respectable in a mid-range unit, and the better series are indistinguishable in service. The series matters more here than the brand.
Worth a second look
Not automatically bad, but historically associated with early failure in hot positions. On a unit sold today it says something about the budget it was built to.
Two caveats. A Japanese bulk capacitor says nothing about the secondary capacitors, which sit closer to the heat and fail sooner — a unit can advertise "Japanese capacitors" truthfully while using them in one position. And counterfeits of the well-known series do circulate on the cheapest units, so on an implausibly inexpensive supply the printed name proves nothing.
, and behaviour can only be measured. No datasheet reports them honestly and no photograph reveals them, which is why the on this site caps a rating when no laboratory data exists — however good the architecture looks on paper.
In practice
The short version, in the order the decisions happen.
Reference
Every term the site uses, in plain language, with a note on why it should matter to you. These same definitions appear as tooltips wherever the term is used.