ATX 3.0 vs ATX 3.1: Do You Need a New PSU for the RTX 50 Series?

Almost every argument about this topic goes wrong in the same place, so it is worth being precise before anything else. ATX is a design guide published by Intel and addressed to power supply manufacturers. Intel's own description is that it “is mostly commonly referred to as the ATX specification for power supplies”. PCIe CEM is a separate specification published by PCI-SIG, addressed to the card side. Statements about one are constantly attributed to the other.

Intel's revision ladder is also confusing by accident. The document is number 336521, and the marketing names do not appear in its title. Revisions 2.0 and 2.01 are titled “ATX Version 3.0 Multi Rail…”; from revision 2.1 Intel dropped the “.0” from the title entirely, and the tie to the name “ATX 3.1” appears inside the text: “This document aligns to ATX Specification Version 3.1.” Revision 2.1 is dated September 2023, and the current revision 2.1a November 2023.

The useful part: Intel publishes all of these openly on its documentation site, each revision at its own address, classified public. That means the two versions can be compared directly rather than argued about — which is what the rest of this article does.

The substance of ATX 3.0 was the arrival of an entire chapter on add-in card power behaviour. Its revision history records “Added ALL of Section 3 — PCIe* Add-in Card Consideration” alongside the new 12VHPWR connector.

Two requirements matter to a buyer. On the card side, Intel's Table 3-1 caps how far an add-in card may exceed its sustained power over a given interval — for cards above 75 W the ratio reaches 3x at intervals up to 100 microseconds. This is the origin of the widely repeated “GPUs can spike to three times their rated power”. It is true, with a qualifier the folklore drops: 3x is the limit at the shortest interval, and the permitted ratio falls as the interval lengthens.

On the supply side, Table 3-3 sets what a power supply must ride out: excursions to 200% of its rated size at 100 microseconds, 180% at 1 millisecond, 160% at 10 milliseconds, and 120% at 100 milliseconds.

That 200% figure is also conditional, and this is routinely dropped. The table has two columns. The higher ladder applies to supplies above 450 W that carry the 16-pin connector; smaller units without it are held to a lower bar. Intel explains the split directly: “The largest power excursion coming from PCIe Add-in cards comes from the 12V-2x6 connector, therefore the table is split into two different power levels where it is expected that power supplies greater than 450 watts will have at least one 12V-2x6 connector.” A 400 W ATX 3.x unit was never required to hit 200%.

The change is in one revision-history entry, and it is entirely about the connector: “The 12VHPWR connector's PCB Header's internal pin lengths have been modified. Therefore, the connector has been given a new name. 12VHPWR connector name has been changed throughout the document to ‘12V-2x6’. The ‘Cable Plug’ side of the connector has not changed and is compatible with new PCB Header connector definition.”

Now the part that is worth the price of the whole article. The transient requirement did not move. Table 3-3 is numerically identical in ATX 3.0 revision 2.0, ATX 3.0 revision 2.01, and ATX 3.1 revision 2.1a — the same 200 / 180 / 160 / 120 ladder against the same intervals. Both popular claims are therefore false: ATX 3.1 did not tighten the excursion requirement, and it did not relax it.

There is a real duty-cycle relaxation people point at, and it is worth knowing where it actually happened: inside ATX 3.0, at revision 2.01, roughly eight months before ATX 3.1 existed. Attributing it to 3.1 gets both the version and the date wrong.

So on the power-delivery side, a compliant ATX 3.0 supply and a compliant ATX 3.1 supply are held to the same standard. The difference between them is which connector definition they were built against — and even there, Intel says the cable did not change.

This is the question the article's title asks, so it deserves a direct answer from the vendor rather than an inference. NVIDIA's RTX 5090 product page gives a wattage and one sentence of qualification: “Minimum is based on a PC configured with a Ryzen 9 9950X processor. Recommend PCIe CEM 5.1 compliant PSU.”

Three things follow. NVIDIA cites PCI-SIG's CEM specification, not Intel's ATX design guide — the strings “ATX 3.0” and “ATX 3.1” do not appear. It recommends rather than requires. And its wattage figures are stated against a named test configuration, which is why they are best read as a floor for a comparable build rather than a universal answer.

The in-box documentation is even quieter. The RTX 5090 Quick Start Guide does not mention ATX, CEM, transients, spikes, 12VHPWR or 12V-2x6 anywhere. It does tell you to use the included adapter with a Founders Edition card, which means NVIDIA treats the bundled 8-pin adapter path as a fully supported configuration — not a compromise to be upgraded away from.

And the finding that answers the title most directly is an absence: no Intel or NVIDIA document we could reach states that an ATX 3.0 unit is insufficient for an RTX 50 card. Not a failure to find one — we went looking in the design guide, the product pages and the in-box guides, and the claim is simply not made by anyone in a position to make it.

No official document anywhere sets a generational rule for this, and it is worth saying why: Intel's document is a design guide addressed to power supply manufacturers, not an upgrade guide addressed to you. Nobody official publishes “replace your PSU when X”. What follows is our reasoning from the requirements above, not a quotation.

Replace it for wattage. This is the reason that actually applies most often and has nothing to do with which ATX revision is printed on the box. Our per-card pages and the PSU calculator will give you a figure for your own combination of parts.

Replace it for cabling. A card that wants two or three independent 8-pin feeds, or a native 16-pin lead, is a legitimate reason to buy — an adapter chain adds joins to the highest-current path in the machine, and a supply without enough separate cables forces daisy-chaining.

Replace it for age or quality. Capacitors age, and a supply sized with no margin when it was new has less as it gets older. A unit of unknown provenance running near its limit is worth replacing regardless of any specification revision.

Do not replace it merely because it says 3.0 rather than 3.1. On the evidence above, that difference is a connector-header revision on the graphics-card side plus a rename — with an identical transient requirement on the supply side.

There is no Intel certification programme behind that phrase. The document is a design guide, and its labelling and marking chapter is classified RECOMMENDED rather than required — the heading itself reads “Labeling and Marking - RECOMMENDED”. Even the connector labelling clause is written as a recommendation: hard-wired 12V-2x6 harnesses “shall be labeled indicating the maximum power supported”, inside a chapter Intel does not make mandatory.

So a badge on a box is a manufacturer's self-declaration that its design follows the guide. That is not worthless — a reputable manufacturer's claim carries real weight — but it is not a test result from an independent body, and it should not be read as one.