What PSU Do RTX 50-Series Cards Need? 5070 to 5090 (2026)

Buy NVIDIA’s number for your card, then check two things. The required system power runs 650W (RTX 5070), 750W (5070 Ti), 850W (5080), 1000W (5090) — and the two things that move it are which board you bought and which processor sits next to it. A factory-overclocked card or a 200W-plus CPU costs a full power-supply tier at every step of the range, and on the 5090 our own itemized math lands above NVIDIA’s floor, which happens nowhere else in the lineup.

Every card’s actual power specification

Two details matter more than the headline row. The TGP is the card; the required figure is the system, already carrying an allowance for processor, drives, and fans. And the connector count climbs with the tier — an older supply with two spare 8-pins runs a 5070 out of the box, needs a third cable for a 5080, and is realistically out of its depth on a 5090, where NVIDIA also recommends a PCIe CEM 5.1 compliant unit rather than just a wattage.

The sizing method, stated once

Every table below sums real component draw the way our PSU wattage calculator does; the per-component coefficients and their sources are published there. The reference build — two RAM sticks, one M.2 drive, four fans, an AIO — adds 93W on top of card and processor. “Smallest that works” is peak draw plus a 20% margin, rounded up to a wattage actually sold; one size above that is the quieter, longer-lived choice. Whether a processor limits your frame rate rather than your power budget is a different question — the PC bottleneck calculator answers that one.

RTX 5070: 650W, honestly earned

The mainstream gaming row lands exactly on NVIDIA’s 650W — a useful cross-check, though NVIDIA does not publish the build behind its figure. Go to 750W for a flagship CPU (the 250W row peaks at 91% load on a 650W unit, past the point where a supply gets loud and protection-trippy) or if a 5070 Ti upgrade is on the table, since the Ti asks for 750W anyway. A genuinely lean build — 65W chip, one drive, no AIO — sits on 550W by the arithmetic, but quality ATX 3.x units below 650W are thin on the ground, so if the two shelves are twenty dollars apart, take the 650W. Weighing the Ti now? The RTX 5070 vs 5070 Ti comparison covers what the extra 50W buys; for the AMD side, RTX 5070 vs RX 9070.

RTX 5080: which 5080 did you buy?

This is the tier where the board variant decides the answer. A Founders-class card is 360W; factory-overclocked models are sized nearer ~420W under sustained load, and that 60W is worth an entire supply tier at the top of the CPU range:

Read across any row: the overclocked board costs you a tier on its own — same silicon, different power target. Our arithmetic for a Founders-class card with a mainstream CPU lands at 750W, below NVIDIA’s 850W; that is not a gap to exploit. Vendor minimums are worst-case by design — read the 850W as covering the bottom of the table rather than the middle, and buy it.

What our own overclocked 5080 actually drew

Our reference card is the overclocked kind, and we logged it: across roughly 26 logged minutes of gaming, captured in two segments, on the ASUS ROG Astral RTX 5080 OC — 308 HWiNFO64 samples on the hand-tuned profile — board power averaged 261W and peaked at 326W.

Both numbers sit comfortably under the ~420W the calculator uses for an overclocked board — which is the point. You size a supply for the worst case it must survive, not for the average it will usually see.

RTX 5090: the one card where the math flips

On the 5070 and 5080 our arithmetic comes out below NVIDIA’s minimum, and the honest advice is to buy the vendor number anyway. The 5090 inverts that. NVIDIA lists 1000W and states the assumption behind it: a system built around a Ryzen 9 9950X. Run the itemized math on exactly that pairing:

Read the third row: on the exact processor NVIDIA cites, the rule says 1200W where they say 1000W. A 1000W unit is not wrong — 868W peak is 87% load, and it will run — but it sits above the 80% line where efficiency sags, and it leaves nothing for an overclocked board or a future card. 1000W is the floor. 1200W is the buy.

The 5090 is also where the connector stops being a preference. Four 8-pin adapters stacked behind a card is where cable management stops being cosmetic, so most 5090 builds end up on a native 12V-2x6 cable — and NVIDIA’s CEM 5.1 recommendation is a validation standard, not a wattage: a 1200W unit that predates it is a worse pairing than a compliant 1000W one. Our PSU tier list marks which current units carry a native 12V-2x6 header.

If you are buying the whole tower rather than the card, the best RTX 5090 prebuilt guide compares three current 9800X3D systems by the exact power supply, memory, cooling, seller, and price details each listing exposes.

Transient spikes, honestly

Graphics cards draw in bursts, and this is where PSU advice usually turns into “buy double.” It should not. Intel’s ATX 3.0 design guide requires a compliant unit to survive excursions to 200% of its rated wattage for 100 microseconds, with lower brackets for longer events — but only where both conditions hold: the unit is above 450W and ships a native 12VHPWR / 12V-2x6 connector. A compliant modern unit at the sizes in these tables already has spike headroom in the standard; size for sustained peak and stop padding. An older ATX 2.x supply was never designed against those brackets — leave more room, and the bigger the card, the less proportional margin a spike has above your peak.

Whichever connector you use, seat it fully until it clicks and avoid a sharp bend at the plug. The current 12V-2x6 design shortens the sense pins precisely so a partially seated plug fails to deliver full power instead of making high-resistance contact.

Undervolting: headroom you earn, not money you save

Curve undervolting cuts real watts at every tier — our results database records roughly 50W back on mid-range cards and around 100W on flagship-class boards, and the published range for the 5080 specifically is 30–70W depending on silicon. On the 5090’s 200W-CPU row, 100W off moves a 1000W unit from 87% load back to a comfortable 77%.

The honest caveat: you cannot bank it in advance. Silicon varies, and a supply has to cover the case where your card undervolts poorly. Buy for the stock figure, tune afterwards with the GPU undervolting guide, and treat the saving as headroom you gained.

If you already own the power supply

Check the label on the unit, not the case badge — most supplies report nothing to Windows, so the label’s continuous 12V rating is the number that counts, not the peak sticker. On a multi-rail unit, follow the manual’s rail distribution and give the card its own cables instead of daisy-chaining one pigtail. And a supply that is merely old is not automatically a problem, but one that is failing looks exactly like one that is undersized — random shutdowns under load, reboots in games. PSU failure symptoms covers telling those apart before you spend money on the wrong fix.

Once the wattage is settled, the which unit question is what the PSU tier list answers, with every placement sourced; for what each card buys you in frames, see the GPU hierarchy.

Sources

• NVIDIA GeForce RTX 5070 / 5070 Ti official specifications — TGP, required system power, and connectors — RTX 5070 family specifications, verified live 2026-07-25.
• NVIDIA GeForce RTX 5080 official specifications — RTX 5080 specifications, verified live 2026-07-25.
• NVIDIA GeForce RTX 5090 official specifications — including the stated Ryzen 9 9950X basis for the 1000W figure and the PCIe CEM 5.1 recommendation — RTX 5090, verified live 2026-07-25.
• Intel ATX 3.0 desktop power supply design guide — the excursion brackets and the wattage/connector conditions that qualify a unit (Intel ATX 3.0 PSU power excursion).
• The ~420W (5080 OC) and ~660W (5090 OC) sizing figures are this site’s own conservative estimates from the GPU list in the PSU calculator — sizing peaks, not measured gaming draws.
• The 261W average / 326W peak figures are first-party telemetry: 308 HWiNFO64 samples across roughly 26 logged minutes (two segments) on the hand-tuned Astral 5080 profile, charted above; the raw CSV ships with the Astral bench dataset.
• The 30–70W (5080) undervolt range and the ~50W / ~100W tier figures are the published community numbers recorded in the undervolt results database.
• Component draw coefficients and sizing rules are this site’s own published figures, listed in full in the PSU calculator methodology.