The standard advice is to add up your components’ rated power, add a hundred watts, and buy that. It produces an answer that is usually too large, occasionally dangerously too small, and wrong for an interesting reason.
Why adding up TDP overshoots
A processor’s rated power is what it draws under a sustained all-core load — a stress test, a render, a compile. Games do not do that. A game loads a few threads heavily and leaves the rest idle, so a chip rated at 250 W typically draws 120 to 160 W while gaming.
Build a budget from stress-test figures and you size the supply for a situation your machine will almost never be in. This is how people end up with a 1200 W unit feeding a system that draws 500 W in practice.
There is a real cost to that beyond the money. Power supplies are most efficient around half load, and a unit running at 20% of capacity sits in a less efficient part of its curve. Oversizing does not harm the hardware; it just wastes money twice.
Why the maths can also undershoot
Graphics cards do not draw power smoothly. They draw it in transient spikes — bursts lasting a fraction of a millisecond that can reach two or three times the card’s rated power.
A 350 W card can momentarily pull above 700 W. The average over any sensible window is still around 350, so a wattmeter at the wall will never show it, but the supply has to survive it.
What happens when it cannot is the thing that makes this worth understanding: the supply’s protection circuitry sees an overcurrent condition and shuts down. The machine turns off instantly, mid-game, with no warning and no error. People reasonably assume the graphics card is faulty — the shutdown happens under graphics load — when the supply is the component that failed.
This is why a system drawing 500 W can be unstable on a 600 W supply while running fine on a good 750 W one. The headroom is not for the average. It is for the spikes.
What actually determines the number
The graphics card dominates. On most gaming builds it is more than half the total and essentially all of the transient behaviour.
The processor contributes less than its rating suggests, for the reason above — unless you also render or compile, in which case the stress-test figure is the relevant one and you should size for it.
Everything else is noise. Drives, fans, memory and peripherals together are typically 30 to 60 W. Elaborate per-component calculators that let you specify eight case fans are adding precision to a rounding error.
The supply’s own quality matters more than its label. A well-built 650 W unit handles transients better than a cheap 850 W one. Wattage is not a quality rating, and the units most likely to shut down under spikes are the ones that advertise a high number cheaply.
A sensible way to size it
- Take the graphics card’s rated board power.
- Add a realistic gaming figure for the processor — roughly 60% of its rating for a modern high-core-count chip, closer to its full rating for an older or lower-core one.
- Add 50 W for everything else.
- Add headroom for transients: roughly 65% on top for a small build, 50% for a mid-range one, 35% for a large one. Larger builds need proportionally less because the spikes are a smaller fraction of an already-large total.
- Round up to the next standard size.
That is what our PSU calculator does, and it is deliberately less alarming than most. An earlier version of it recommended 1200 W for a build that genuinely needed 750, because it used stress-test processor figures. That was wrong, and the current version uses gaming-realistic draw with tapering margin.
Connectors matter as much as watts
Worth checking separately, because the wattage can be right and the build still impossible:
- 12VHPWR / 12V-2x6 — current high-end NVIDIA cards use it. An adapter from four 8-pin cables works, but a supply with the connector native is tidier and less failure-prone.
- Number of PCIe cables — some cards need three 8-pin connectors, and some supplies provide two.
- Daisy-chained cables. Many supplies ship cables with two connectors on one lead. On a high-draw card, use separate cables for each connector rather than one cable doing double duty.
What not to pay for
Efficiency ratings above Gold. Platinum and Titanium cost meaningfully more and save a few watts. The payback period on the price difference is usually measured in years.
Capacity you will never use. If the calculation says 750, a 1000 W unit buys you upgrade headroom and nothing else. That is a legitimate reason to buy one, but it is a decision about your next graphics card, not about this build’s stability.
Wattage from an unknown brand. The number on the box is the cheapest thing to print. A supply from a manufacturer with a reputation to protect, at the capacity you actually need, is a better purchase than a larger number from one without.