PSU sizing: 12VHPWR, transient spikes and what 80 PLUS means
How to size a power supply for a GPU with millisecond spikes, why the wattage label is not the whole story, and what to check before buying.
- Build
- 26300.9457
- Edition
- Pro (26H2)
- Architecture
- x64
- Support
- Documentation (E1)
- Evidence
- E1Official documentation
- Verified
- 2026-10-29
- Page type
- Concept
The short answer
Section titled “The short answer”Size the PSU for transient spikes, not for the average wattage on a spec sheet:
- Add up sustained power: CPU package + GPU board power + ~50–80 W for the rest.
- Add headroom for millisecond spikes — modern GPUs can pull well above their rated power for a few milliseconds.
- Check the connectors the GPU needs (12VHPWR/12V-2×6, or 2–3× 8-pin).
- Buy quality, not just watts: protection circuits, a reputable platform, a warranty you can use.
Under-sizing does not usually present as “it will not boot”. It presents as random reboots, storage corruption, or a GPU that drops off the bus under load — symptoms that look like software bugs.
The numbers that matter
Section titled “The numbers that matter”| Component | Sustained | Notes |
|---|---|---|
| CPU | package power (PL1/PL2) | PL2 can be far above the “TDP” |
| GPU | board power | spikes higher than the rated number |
| Motherboard/RAM/SSD/fans | 30–80 W | NVMe and RGB add up |
| Transient headroom | +20–30% | covers GPU spikes |
A 450 W system does not need a 1200 W PSU, but it does need one that can survive a 600 W microsecond spike without tripping.
12VHPWR: what changed
Section titled “12VHPWR: what changed”New high-end GPUs use a 16-pin connector (12VHPWR, revised as 12V-2×6). Practical rules:
- Use the cable that came with the PSU, or a vendor-approved cable for that exact model. Pin ratings are not interchangeable across brands.
- Fully seat the connector; partial insertion is the documented cause of melted connectors.
- Avoid tight bends within ~35 mm of the connector.
- A 12VHPWR adapter from multiple 8-pin cables is acceptable only if the PSU vendor supplies it.
Sizing table by build
Section titled “Sizing table by build”| Build | Typical sustained | Sensible PSU |
|---|---|---|
| Office/mini PC | 100–150 W | 300–400 W (often external brick) |
| Mainstream gaming | 300–400 W | 550–650 W |
| High-end gaming | 450–600 W | 750–850 W |
| Workstation with one big GPU | 500–700 W | 850–1000 W |
| Multi-GPU/compute | 800 W+ | 1200 W+, with proper circuit planning |
These are starting points, not guarantees — a 250 W GPU with a 250 W CPU on the same rail changes the answer.
Check what your machine actually draws
Section titled “Check what your machine actually draws”# rough load indicator, not a wattmeterGet-Counter '\Processor Information(_Total)\% Processor Time' -MaxSamples 5Get-CimInstance Win32_Processor | Select-Object Name,MaxClockSpeed,NumberOfCoresWindows cannot measure wall power. A plug-in wattmeter (or a PSU with monitoring) is the only honest way to know your real draw — and it is what you need before adding a second GPU.
Symptoms of an under-sized or failing PSU
Section titled “Symptoms of an under-sized or failing PSU”| Symptom | Why it points at the PSU |
|---|---|
| Reboots only under GPU load | transient spike trips OCP |
| GPU disappears from Device Manager under load | rail sag |
| Storage errors that move between disks | unstable 5V/3.3V |
| Coil whine that scales with load | inductor/load interaction (usually harmless) |
| Machine will not resume from sleep | standby rail |
Test by removing load (one GPU, fewer drives) and repeating the workload. If the symptom disappears, the PSU or its cabling is the suspect.
Related
Section titled “Related”Lab boundary
Section titled “Lab boundary”E1. Power figures are component-vendor specifications, not measurements made here. The lab has no wattmeter and no discrete GPU, so no power data is published.