Guide

GPU transient spikes, explained

A graphics card's rated board power is an average it is allowed to sustain. For a few millionths to a few thousandths of a second it can pull more. That spike is real. Adding it on top of the wattage, as if the card drew double forever, is how a calculator turns a 300 W GPU into a 1,000 W shopping suggestion.

Two different limits

Intel's ATX power-supply design guide, document 336521, was updated for PCI Express CEM 5.1. It separates what the card may do from what the supply must absorb.

On the card, Table 3-1 allows a card above 75 W to reach 3× its sustained power for intervals of 100 microseconds or less. Between 100 microseconds and 1 second the allowed ratio follows a logarithmic curve in the spec. At 1 second and beyond the card is back to its sustained power. The card also has to stay inside the voltage and current limits of the connector.

On the supply, Table 3-3 splits the test by connector. A unit above 450 W that has a 12V-2x6 connector must support 200% of its label for 100 µs, 180% for 1 ms, 160% for 10 ms, and 120% for 100 ms, at duty cycles of 5%, 8%, 12.5%, and 25%. The same table's other column, for units at or under 450 W and for larger units without that connector, is 150%, 145%, 135%, and 110% at those same times. Continuous power is 100% in both columns.

Read those percentages against the PSU label, not against the GPU. A 1,000 W ATX 3.1 unit is tested to 2,000 W for 100 microseconds. That is the spike budget. You do not also add 2,000 W to the shopping list.

The connector tells the card how much it may take

The 12V-2x6 plug replaced 12VHPWR. Four sense pins advertise the cable's rating. Intel's Table 3-6, following CEM 5.1, maps them like this: both sense pins grounded means up to 600 W sustained after configuration, one open and one grounded means 450 W, the opposite pair means 300 W, and the two pins shorted to each other means 150 W. Open and open means the card should not draw on that cable. The card can draw that cable budget in addition to the slot.

An 8-pin is 150 W and a 6-pin is 75 W under the older PCI-SIG cable ratings, plus up to 75 W from the slot. WattWise checks the rated board power against those budgets and says when a path is tight. It does not invent a spike wattage for each GPU, because the vendors do not publish one.

What to buy, and what not to do

For a card whose reference design offers a PCIe Gen 5 cable, use a native 12V-2x6 lead rated at least as high as the vendor's requirement, on an ATX 3.1 (or ATX 3.0) supply, and seat it until the latch clicks. Adapters exist because NVIDIA ships them. They are a stack of 8-pin plugs, and every plug has to be fully in. A half-seated high-current connector is a thermal problem, not a wattage problem.

Older 8-pin cards, including most Radeon RX 6000 and 7000 reference designs and the Arc cards listed here, do not need a 12V-2x6 lead. They need the plugs the vendor lists, and a continuous wattage that covers the estimate. The lower excursion column still applies if the supply has no 12V-2x6 connector. That is a reason to buy a competent unit, not a reason to double the label.

The scary oscilloscope traces in reviews are usually microseconds wide. The supply's output capacitors and the excursion spec are what those traces are asking for. If a game crashes, the cause might be a spike, a weak unit, a bad cable, or an unstable undervolt. The calculator cannot see your cables. It can refuse to pretend that a 200% microsecond event is a 200% shopping requirement.

Sources: Intel document 336521 and the PCI-SIG power-excursion ECN. The same rules are summarized on the methodology page and on every GPU guide that uses a Gen 5 connector.