Sizing a Power Supply for Transient GPU Spikes

What a transient actually is

Published power figures are averages over a measurement window. Real electrical demand is not smooth: a graphics card’s load changes as fast as its workload does, and the switch from idle to a full compute burst happens in microseconds.

Those brief excursions are considerably higher than the average. They last far too short a time to register on a wall-socket power meter or in monitoring software that samples once a second, which is exactly why they surprise people — every tool you might reach for to check reports the average and misses the spike entirely.

A power supply, however, does not miss them. Its protection circuitry monitors current continuously, and a large enough excursion trips over-current or over-power protection. The supply shuts down to protect itself and the components attached to it, which is the correct behavior and an infuriating one.

The vendor numbers as evidence

You do not have to take the phenomenon on trust, because the card manufacturers publish the implication. NVIDIA’s specification for the RTX 4090 pairs a 450 W total graphics power with a required system power of 850 W. For the RTX 5080 the total graphics power is 360 W.

Consider what 850 W has to cover. A 450 W card plus a processor at, say, 250 W maximum turbo is 700 W of rated draw, leaving 150 W for drives, fans, memory and the board — which is generous but not enormous. The recommendation is not simply the sum of the parts; it embeds the assumption that a supply operating well below its ceiling has the margin to absorb PSU transient spikes without tripping.

Read vendor system-power recommendations as a statement about headroom against PSU transient spikes rather than about total consumption, and the numbers stop looking inflated.

The symptom you will actually see

This failure has a distinctive signature, and recognizing it saves days of misdirected troubleshooting.

• The machine shuts off instantly — no error, no crash dialogue, no blue screen. Power simply stops, as if the plug were pulled.
• It happens under load, often at a scene transition or the start of a demanding sequence rather than during steady heavy use.
• It does not happen in stress tests. Synthetic tests hold a steady high load, which is precisely the condition transients do not occur in.
• Temperatures are fine. Everything logged looks healthy right up to the moment the log ends.
• Capping the frame rate or power limit stops it. This is the strongest single diagnostic.

That last point is worth trying before buying anything. If limiting the card’s power to 80% eliminates the shutdowns entirely, you have confirmed the diagnosis for free and you have a temporary workaround.

Sizing it properly

1. Start from the vendor’s required system power for your graphics card. This is the single most reliable input available.
2. Add your processor’s maximum turbo power if it exceeds what the vendor’s typical-system assumption allowed for — a 250 W processor is above average.
3. Add 50 to 75 W for drives, fans, memory and peripherals in a typical build.
4. Target roughly 50 to 70% loading at your normal operating draw. Supplies are most efficient there and have the most margin above it.
5. Prefer a modern standard with defined excursion tolerance and the correct native connector for your card, rather than adapting older cabling.

On connectors specifically: use the cable the supply shipped with, seated fully. Adapter chains from multiple older connectors are a documented source of trouble on high-power cards, and a partially seated high-current connector is a genuine hazard rather than merely a performance issue.

Why the unit matters more than the number

Two supplies with the same wattage on the label can behave completely differently when handling PSU transient spikes, because the specification that matters is how the protection circuitry is tuned and how much reserve the design holds.

A well-engineered 850 W unit rides out excursions that trip a poorly engineered 1000 W one. So capacity is necessary and not sufficient: look for a supply from a manufacturer with a published, independently reviewed platform, a long warranty as a proxy for the manufacturer’s own confidence, and an efficiency certification that indicates the design was built to a standard rather than to a price.

Buying a larger cheap supply instead of an adequate good one is the most common way people spend money on this problem without solving it.

The checklist

• Use the vendor’s required system power as your starting figure, not the card’s TGP.
• Account for a high-turbo processor separately.
• Aim to run at 50–70% load in normal use.
• Test the diagnosis by capping power to 80% before buying anything.
• Use native cables, fully seated. Avoid adapter chains.
• Choose a reviewed platform with a long warranty over a bigger unknown one.

Approached that way, sizing for transients is a five-minute calculation with one free test attached — and it explains the class of instant shutdowns that no temperature log or stress test will ever reproduce. Our cooling and power supply guide covers selecting the unit itself.