Calculators

AI hardware

PC PSU Wattage Calculator for AI and Multi-GPU Builds

Estimate the power supply size for a local AI workstation, multi-GPU box, or creator rig before buying parts.

When to use this

Use it before sizing a high-power local AI PC

This calculator is for GPU inference, fine-tuning, rendering, batch generation, and multi-GPU experiments where sustained draw and cabling matter more than a generic desktop estimate.

Default result

The server-rendered dual RTX 4090 example estimates 1,199 W sustained draw and recommends a 1,600 W PSU.

Estimate PSU wattage

Start with a build preset, then tune GPU count, CPU watts, accessory load, and reserve assumptions.

GPU load

900 W GPU sustained

NVIDIA GeForce RTX 4090: 450 W TGP per card.

Other system load

299 W

CPU, board, drives, fans, pump, USB, and PCIe accessory allowance.

Cable note

2 GPUs: NVIDIA lists 3x PCIe 8-pin adapter cables or 1x 450 W or greater PCIe Gen 5 cable for the Founders Edition.

Use separate vendor-approved cables for high-power GPUs.

Worked example: dual RTX 4090 AI workstation

The default build uses 2 x RTX 4090 at 450 W each, so GPU TGP is 900 W. At 100% AI load, sustained GPU planning draw stays 900 W.

The rest of the system is 170 W CPU + 60 W motherboard + 2 x 8 W drives + 6 x 3 W fans + 10 W pump + 25 W USB and PCIe allowance = 299 W.

Sustained system draw is 900 W + 299 W = 1,199 W. With 5% aging reserve and a target of keeping sustained load at 80% of PSU capacity, target capacity is 1,199 x 1.05 / 0.80 = 1,573.7 W.

The transient check is 900 W x 1.20 + 299 W = 1,379 W. The larger planning number is 1,573.7 W, so the calculator rounds up to a 1,600 W PSU. Sustained load on that PSU is about 1,199 / 1,600 = 74.9%.

How the AI PSU wattage estimate works

The calculator starts with GPU total graphics power multiplied by GPU count and sustained load percentage. It then adds CPU package watts, motherboard allowance, storage, fans, pump, USB, and PCIe accessory load.

The recommendation is the larger of two checks: sustained watts with aging reserve divided by the target PSU load, or GPU TGP with transient reserve plus the rest of the system. The result is rounded up to a common PSU size.

Why multi-GPU AI builds need more margin

AI workloads can run sustained GPU load longer than gaming bursts, and multiple cards concentrate heat and cable current. A build that technically boots on a smaller PSU may still be noisy, unstable, or uncomfortable for long unattended jobs.

What PSU size for two RTX 4090 GPUs?

With the default 170 W CPU and accessory assumptions, two RTX 4090 cards produce a 1,199 W sustained estimate and a 1,600 W rounded recommendation. A lower-power CPU, power-limited GPUs, or measured draw can reduce the number, but the cabling still needs careful verification.

How to use measured wall power

If you already own the system, measured wall power is useful but includes PSU losses. For planning, use component DC draw when possible, then compare wall measurements under the exact training, inference, rendering, or stress workload you care about.

What to do next

After the wattage screen, verify case clearance, airflow, individual cable ratings, connector type, branch circuit capacity, UPS size, and the exact add-in-card vendor specs before buying a PSU.

Reference data used by the defaults

Topic Reference value Source Date Note
GeForce RTX 5090 reference power 575 W total graphics power; 1000 W required system power; 1x 600 W PCIe Gen 5 cable or 4x PCIe 8-pin adapter cables. NVIDIA RTX 5090 specifications As of June 19, 2026 NVIDIA says power requirements can vary by system configuration and card vendor.
GeForce RTX 5080 reference power 360 W total graphics power; 850 W required system power; 1x 450 W or greater PCIe Gen 5 cable or 3x PCIe 8-pin adapter cables. NVIDIA RTX 5080 specifications As of June 19, 2026 The calculator uses TGP as the GPU sustained planning input, then adds workload and headroom assumptions.
GeForce RTX 4090 reference power 450 W total graphics power; 850 W required system power; 1x 450 W or greater PCIe Gen 5 cable or 3x PCIe 8-pin adapter cables. NVIDIA RTX 4090 specifications As of June 19, 2026 Use the add-in-card vendor page for factory-overclocked boards.
GeForce RTX 4080 SUPER reference power 320 W total graphics power; 750 W required system power; 1x 450 W or greater PCIe Gen 5 cable or 3x PCIe 8-pin adapter cables. NVIDIA RTX 4080 family specifications As of June 19, 2026 The page covers RTX 4080 SUPER and RTX 4080 reference-family values.

GPU board power and connector requirements can differ by add-in-card vendor. Replace defaults with the exact model you will buy.

FAQ

Is this the same as NVIDIA required system power?

No. NVIDIA required system power is a vendor reference for one GPU in a stated system. This calculator rebuilds the estimate from your GPU count, CPU watts, accessories, target load, and reserve assumptions.

Why target 80% PSU load instead of 100%?

A sustained AI workload can run for hours. Keeping the expected sustained load below the full PSU rating leaves thermal, acoustic, transient, cable, and aging margin. You can edit the target if you have a different design rule.

Do AI workloads really use full GPU board power?

Some do and some do not. Training, fine-tuning, batch inference, rendering, and stress tests can approach board power. Light chat inference may sit lower. Use measured wall power if you already have the hardware.

Can I use splitter or daisy-chain GPU power cables?

Do not assume that is safe. Follow the GPU and PSU vendor cabling instructions, use enough dedicated cables, seat high-power connectors fully, and avoid bending cables sharply near the connector.

Does PSU efficiency reduce the wattage I need?

No. PSU efficiency changes wall draw and heat, not the DC wattage your components need. A 1,200 W PSU is rated for DC output; an efficient unit simply pulls less extra power from the wall.

When should I ask a system builder or electrician?

Ask for help when the build uses multiple high-power GPUs, custom cables, rack PSUs, breakers near their limit, international voltage constraints, or sustained unattended workloads.

Decision path

What to do next