Home › Guides › Hardware, power and cost

Sizing a Power Supply for an RTX 4090 Build: ATX 3.x Explained

Updated October 21, 2026 · By the Xingyu Zhitong editors

Power supply advice for high-end GPUs tends to collapse into a single number ("get 1000 W") that is sometimes right for the wrong reasons. A better approach separates three questions: how much sustained power the system draws, how high short spikes go, and whether the connectors and cables are rated for the current involved. This article works through each for a typical RTX 4090 workstation, and explains what the ATX 3.0 and 3.1 standards changed.

Start with what NVIDIA publishes

NVIDIA's RTX 4090 page lists a total graphics power of 450 W for the reference design and a required system power of 850 W. The page states that the 850 W figure assumes a PC built around a Ryzen 9 5900X and that a higher-rated supply may be needed depending on the system. It also lists the power connection: either three PCIe 8-pin cables through the adapter included in the box, or a single PCIe Gen 5 power cable rated at 450 W or greater. The card has 24 GB of GDDR6X on a 384-bit interface. Partner cards may differ from the reference values, so check the specific model's documentation; some factory-overclocked designs have higher power limits.

The 850 W recommendation is a floor for a mainstream build. If your CPU draws more than the reference platform's, or you add many drives, fans and PCIe cards, your own sum may exceed it.

A component-by-component sum

Add up worst-case sustained draw for every component that can run at the same time:

• GPU: 450 W reference board power (check your model's power limit).
• CPU: for an Intel Core i9-14900K, Intel's published specification lists 125 W processor base power and 253 W maximum turbo power (confirm on Intel ARK for your exact SKU). Many motherboards ship with settings that allow sustained power above these values, so reading the real figure for your board matters. AMD's high-end desktop parts have their own published package power limits; use the vendor's spec page.
• Motherboard, RAM, SSDs, fans, pumps, USB devices: 50 to 100 W is a common allowance for a well-equipped build (an estimate, not a specification).

Example (all values from the list above): 450 + 253 + 75 = 778 W worst-case sustained, when both CPU and GPU are fully loaded. Add headroom so the PSU runs below its maximum in normal use; many engineers aim for the typical load to be roughly 50 to 70 percent of the rating, where efficiency is high and fans stay quiet. 778 W ÷ 0.8 ≈ 970 W, so a 1000 W unit is a sensible choice for this build, while 850 W would run close to the limit under a combined stress test. If your workload loads only the GPU (most LLM inference does), the CPU rarely reaches its maximum and the real draw is lower, around 450 W plus a modest CPU and system load. You can model your own parts in the RTX 4090 and i9-14900K PSU calculator.

Sustained load versus transients

Modern GPUs can change their power draw in microseconds. Short spikes above the average board power are normal and are the reason engineers add margin beyond the simple sum. Two details matter here.

First, PSU protection circuits (over-power or over-current protection) can trip on a spike even if the average draw is within rating. Systems that randomly shut down only under GPU load, especially at the start of a render or when a model loads, are classic symptoms.

Second, the ATX 3.0 revision of Intel's power supply design guide was written with this behavior in mind. Intel's ATX 3.0 and 3.1 guides define requirements for tolerating short-duration power excursions from GPUs well above the PSU's rated power. I have not reproduced the exact excursion percentages and durations here because they vary by PSU wattage class and revision; consult the current Intel ATX design guide or the PSU manufacturer's compliance statement for the exact table. The practical point is that a PSU marked ATX 3.0 or 3.1 compliant has been designed and tested for this behavior, while an older ATX 2.x unit of the same wattage may have thinner margin against spikes.

The connector: 12VHPWR and 12V-2x6

The 16-pin connector on RTX 40 series cards (12VHPWR, defined in PCIe CEM 5.0 and ATX 3.0) carries up to 600 W through twelve power pins and four sideband signal pins. The sideband pins let the PSU advertise how much power the cable can safely deliver. With ATX 3.1 and PCIe CEM 5.1, the connector was renamed 12V-2x6 and the pin geometry on the receptacle side was adjusted. Secondary technical coverage reports changes to pin lengths so that the sense pins connect later than the power pins, and shorter hold-up time requirements; confirm details in the official documents if they matter for your design. The plug on the cable looks the same, and the new and old versions are intended to be mutually compatible.

What this means in practice:

• Seat the connector fully. Partial insertion reduces contact area and increases heating. Many reported failures of 16-pin connectors have been attributed to incomplete seating or sharply bent cables. This is a field-experience point rather than a specification, so treat it as a precaution.
• Do not bend the cable right at the connector. Leave several centimeters of straight run before any bend. If your case is narrow, a right-angle adapter or a different case may be safer than forcing the side panel closed.
• Prefer the native cable from a modern PSU over the three-way 8-pin adapter if you have the option, since it removes an adapter and some connections. NVIDIA's own page lists both options as valid.
• Do not daisy-chain or split. Use separate PSU-side connectors for each 8-pin lead of the adapter, as the instructions say.

Running more than one GPU

If you plan a two-GPU workstation for LLM work, the arithmetic changes quickly: two 450 W cards plus a CPU reach about 1.2 kW before headroom, which exceeds most household PSU choices and approaches the limit of a standard 15 A, 120 V circuit (1,800 W nominal, with continuous loads conventionally held to 80 percent, or 1,440 W). Electrical codes vary by country, so check your local rules, and ask a qualified electrician if a dedicated circuit is needed. Many people instead cap the power limit of each card in software (reducing board power by 20 to 30 percent often costs only a modest speed loss in inference, though the effect depends on the workload) or run a larger model with fewer cards. Remember also that the RTX 4090 does not support NVLink, so cards communicate over PCIe; see NVLink versus PCIe before assuming two cards behave like one bigger one.

Diagnosing a marginal supply

If you already own the parts and suspect the supply, work through symptoms before buying. A hard shutdown with no warning, only when the GPU ramps up, points to over-current protection and transient margin. Reboots under combined CPU and GPU stress tests point to total capacity or a failing unit. Coil whine that appears only under load is usually harmless and varies from sample to sample. Visible discoloration at the connector is a safety issue, not a tuning issue. Logging board power with the vendor's monitoring tool while you reproduce the failure tells you whether the card is sitting at its limit; if it is, lowering the power limit by 10 percent as a test often separates a supply problem from a cooling or driver problem. Change one thing at a time and write down the result, because intermittent power faults are easy to misattribute.

Efficiency ratings and what they tell you

80 PLUS ratings (Bronze through Titanium) describe how much of the wall power reaches the components at specific load points. A Gold unit wastes less as heat than a Bronze one, which matters mostly for noise, heat in the case and electricity cost. The difference over a year is small for a home desktop: if your system averages 400 W at the wall for 8 hours a day, a three-percentage-point efficiency gain saves roughly 12 W, about 35 kWh a year, which is a few dollars at typical tariffs. Choose by build quality, warranty length, protections and ATX 3.x compliance first; efficiency tier second. For the larger-scale view of power cost, see PUE and electricity cost.

A checklist for choosing and installing

1. Sum worst-case component power using vendor specifications. Add 20 to 30 percent headroom.
2. Choose a PSU at or above that, preferably ATX 3.0 or 3.1 compliant with a native 16-pin cable if you use an RTX 4090 class card.
3. Check case clearance for the card, the connector and the cable bend radius before buying.
4. Verify that the PSU has enough separate PCIe 8-pin outputs if you use the adapter.
5. Seat connectors fully; inspect after the first few hours under load, and periodically thereafter, for discoloration or heat.
6. Use a wall power meter during your actual workload to measure real consumption rather than relying on estimates.
7. Consider setting a GPU power limit if your PSU is marginal or the case runs hot.

When to ask a professional

If a connector shows melting, discoloration or a burning smell, stop using the system and contact the card or PSU manufacturer about warranty and safety. For wiring your home or office for heavier loads, hire a licensed electrician. This article is general information and does not replace the manufacturer's safety instructions.

Sources

• NVIDIA GeForce RTX 4090 product page
• PCI-SIG (publisher of the PCIe CEM specification that defines the 16-pin connector)
• Intel ARK product specifications (look up your CPU model)
• NVIDIA DGX B200 page (for comparison with data-center system power)

Figures are estimates based on published model configurations and vendor specifications. Real memory use, speed and power draw depend on the software stack, drivers, settings and workload, so test on your own hardware before you buy. Spotted a mistake? Email huhuwen09@gmail.com with the page address and we will check and correct it.