How many watts of PSU do I need for an RTX 4090 or RTX 5090?

For a single RTX 4090 (450 W) with a mainstream CPU and typical peripherals, the component load is around 720 W, so aim for a 1000 W minimum and a 1200 W unit for clean headroom. An RTX 5090 (575 W) pushes the same build to roughly 850 W of load — budget a 1200 W ATX 3.0 PSU.

What size power supply do I need for a dual-GPU (2× 3090) machine-learning rig?

Two RTX 3090s draw 700 W on their own. With a high-core-count CPU (~280 W) and a heavy peripheral load (~160 W) the component load is about 1140 W, which needs a 1600 W PSU. An 850 W unit is not enough — the cards alone exceed it under load.

Do I need an ATX 3.0 power supply for a modern AI GPU?

For high-end cards using the 12VHPWR / 12V-2×6 connector (RTX 40/50 series, RTX 6000 Ada, H100) an ATX 3.0 unit is strongly advised. These GPUs spike to roughly twice their rated power for sub-millisecond windows, and ATX 3.0 PSUs are certified to ride out those transients without tripping.

How much headroom should a PSU have above the GPU + CPU load?

This calculator adds 40% (a ×1.4 factor). That covers GPU transient spikes, CPU boost above base TDP, and keeps the unit near its 50–60% load sweet spot where 80 PLUS efficiency and lifespan are best. Sizing a PSU to run flat-out at 95% load is a common mistake that shortens its life.

How much electricity does a local LLM inference rig use per hour?

It depends on load. A 720 W component load at 80 PLUS Gold draws about 800 W from the wall at full tilt — 0.8 kWh per hour, or roughly Rs 36 at Rs 45/kWh. Inference idles far lower than training, so use the load-factor slider to model realistic usage rather than assuming 100%.

Why is the recommended PSU higher than the wall draw?

They answer two different questions. Wall draw is the AC power you actually consume (component load ÷ PSU efficiency). The recommended PSU is a DC delivery rating that must survive transient spikes and boost, so it includes 40% headroom on top of the steady load — that is why it is larger than both the component load and the wall draw.

Are the GPU and CPU power figures official?

Yes. GPU numbers are the manufacturers' published Total Graphics Power / board TDP from NVIDIA and AMD spec pages. CPU numbers use the vendor's Maximum Turbo Power (Intel) or PPT (AMD) where published, because a PSU must cover boost, not just base TDP. The table was last verified on 2026-07-16.

Can I add a card that isn't in the list?

The GPU list is a curated set of ~23 cards actually used for AI and LLM work. For the CPU there is a Custom TDP option. If you need a GPU that isn't listed, use a listed card with the same board power as a stand-in, or email me its official TDP and I'll add it once verified.

AI hardware · Power

AI GPU Power Supply (PSU) Wattage Calculator

Size the PSU for a single- or multi-GPU AI, local-LLM or Stable-Diffusion rig. Enter your GPUs and CPU to get the recommended wattage, PCIe connectors, ATX 3.0 requirement, wall draw and running cost — using official NVIDIA and AMD power specs. No signup, sources cited.

By Induwara Ashinsana— Executive Director, Ryzera TechnologiesUpdated Jul 16, 2026

Size your PSUGPU + CPU + headroom

NVIDIA / AMD TDP verified

Quick rigs

GPU model

Number of GPUs

CPU

CPU sockets

Storage & peripherals

PSU efficiency (80 PLUS)

GPU overclocked / power limit raised

Adds 10% to the GPU load.

Electricity rate (LKR/kWh)

CEB rates are blocked — Rs 45 is a mid-block estimate.

Load factor for cost: 100%

100% = flat-out training. Inference idles lower.

Recommended PSU

1,200 W

Min 1,008 W · runs at 60% load

Component load (DC)

720 W

GPU 450 W + CPU 170 W + base 100 W

Wall draw (AC)

800 W

At 90% PSU efficiency

Running cost

Rs 36/h

0.8 kWh/h · Rs 288/8h run

ATX 3.0 recommended. The NVIDIA RTX 4090 draws large sub-millisecond transient spikes. An ATX 3.0 / PCIe 5.0 PSU is rated to ride these out; older ATX 2.x units may trip their protection.

Confirm cables: 1× NVIDIA RTX 4090 need 1× 12VHPWR (12V-2x6). Check your PSU actually ships that many native connectors — daisy-chaining or adapters are a common failure point.

Power breakdown

GPU load (1 x 450 W)	450 W
CPU load (170 W)	170 W
Baseline (board, RAM, storage, fans)	100 W
Component DC load	720 W
× 1.4 headroom (minimum PSU)	1,008 W
Recommended PSU	1,200 W

Cross-check: sizing independently for a 60% load target (PSU-vendor sweet spot) gives 1,200 W — the same class, confirming the headroom method.

GPU/CPU power figures are the manufacturers' published TGP/TDP (NVIDIA, AMD, Intel ARK). Headroom follows the Intel ATX 3.0 design guide and 80 PLUS efficiency thresholds. Full source links are listed below the calculator.

How it works

The calculator builds up the total DC power your components draw, adds a safety headroom, and rounds up to the nearest standard retail PSU size. Every input maps to an official specification, and the arithmetic is fully deterministic — nothing is sent to a server.

GPU load. gpuLoad = TDP × quantity. GPU TDPs are the manufacturers' Total Graphics Power figures (e.g. RTX 4090 = 450 W, RTX 5090 = 575 W, RTX 3090 = 350 W, A100 PCIe = 300 W). Turning on the overclock toggle multiplies this by 1.10 to model a raised board power limit.
CPU load.The CPU's Maximum Turbo Power (Intel) or PPT (AMD) is used where published, since a PSU must cover boost, not just base TDP. Dual-socket server selections double it.
Baseline. A fixed allowance for the motherboard, RAM, storage, fans and pumps: Light 60 W, Typical 100 W, Heavy 160 W.
Component DC load. componentLoad = gpuLoad + cpuLoad + baseline.
Headroom. minimum = componentLoad × 1.4. The 40% covers (a) transient GPU spikes up to roughly twice rated power for sub-millisecond windows per the Intel ATX 3.0 / PCIe 5.0 design guide, (b) CPU boost, and (c) keeping the PSU near its 50–60% efficiency sweet spot as PSU vendors (Seasonic, Corsair, be quiet!) advise.
Round up. The minimum is rounded up to a standard retail size — 550, 650, 750, 850, 1000, 1200, 1300 or 1600 W. If the minimum exceeds 1600 W the tool flags that no single consumer PSU is enough.
Wall draw. wallWatts = componentLoad ÷ efficiency, where efficiency comes from the 80 PLUS tier (Bronze ≈ 85%, Gold ≈ 90%, Platinum ≈ 92%, Titanium ≈ 94% at ~50% load, 230 V).
Running cost. kWh/h = (wallWatts × loadFactor) ÷ 1000, then multiplied by your electricity rate. The load-factor slider models the difference between flat-out training (100%) and lighter inference.
Connectors & ATX 3.0.Each card's published PCIe power connectors are summed across the rig, and an ATX 3.0 unit is recommended when any selected card is a high-end, transient-prone GPU on the 12VHPWR / 12V-2×6 connector.

As an independent check, the tool also sizes the PSU a second way — targeting a 60% load point, the PSU-vendor efficiency sweet spot (componentLoad ÷ 0.6). For most builds both methods land on the same PSU class, which is shown as a cross-check line under the breakdown table.

Worked examples

Single RTX 4090 creator / inference PC

1× RTX 4090 · Ryzen 9 7950X · Typical · 80 PLUS Gold

GPU load: 450 W × 1 = 450 W
CPU load: 170 W
Baseline (Typical): 100 W
Component DC load: 450 + 170 + 100 = 720 W
Minimum PSU: 720 × 1.4 = 1008 W → round up → 1200 W
Wall draw: 720 ÷ 0.90 = 800 W
Cost @ Rs 45/kWh, 100% load: 0.80 kWh/h → Rs 36/hour
Connectors: 1× 12VHPWR · ATX 3.0 recommended

Dual RTX 3090 local-LLM training rig

2× RTX 3090 · Threadripper class · Heavy · 80 PLUS Titanium

GPU load: 350 W × 2 = 700 W
CPU load: 280 W
Baseline (Heavy): 160 W
Component DC load: 700 + 280 + 160 = 1140 W
Minimum PSU: 1140 × 1.4 = 1596 W → round up → 1600 W
Wall draw: 1140 ÷ 0.94 = 1213 W
8-hour training run @ Rs 45/kWh: 1.213 kW × 8 h × 45 ≈ Rs 437
Connectors: 6× 8-pin PCIe — confirm your PSU has them
An 850 W PSU is insufficient — the cards alone draw 700 W and total load is 1140 W

Edge case — overclocked RTX 4070 build

1× RTX 4070 (OC) · Ryzen 5 7600X · Typical · 80 PLUS Gold

GPU base load: 200 W × 1 = 200 W
Overclock adder (+10%): 200 × 1.10 = 220 W
CPU load: 105 W
Baseline (Typical): 100 W
Component DC load: 220 + 105 + 100 = 425 W
Minimum PSU: 425 × 1.4 = 595 W → round up → 650 W
Wall draw: 425 ÷ 0.90 = 472 W
No ATX 3.0 flag — the RTX 4070 is not transient-prone

Frequently asked questions

Sources & references

The GPU/CPU TDP table, 80 PLUS thresholds and ATX 3.0 headroom rationale were last cross-checked against these sources on 2026-07-16. Wattage recommendations are a certification-tier guide, not a specific product endorsement.

Related tools

LiveAI

LLM Serving Sizing Calc

Estimate how many concurrent requests one GPU can serve, its sustainable QPS and per-request latency, and the GPU replicas you need for a target load — concurrent users, QPS, or requests/minute. Tied together by Little's Law so concurrency, throughput and latency stay consistent. Free, sources cited.

Open tool

LiveAI

AI Inference Speed Calculator

Estimate how fast an LLM will run on a given GPU — decode tokens/second, prefill throughput, time-to-first-token, and total generation time — from model size, quantization, and published memory bandwidth. Formulas and GPU specs cited.

Open tool

LiveAI

LLM VRAM Calculator

Estimate the GPU VRAM needed to run or fine-tune any open LLM (Llama 3, Mistral, Qwen, Gemma, DeepSeek) at a given precision, context, and batch size — and check whether it fits your GPU. Formulas cited, runs in your browser.

Open tool

Rate this tool

Be the first to rate

Comments & feedback

Spotted a bug or want an improvement? Tell us — our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a bug, edge case, or want a GPU added to the list?

Email me at [email protected] — most fixes ship within 24 hours.