Power, Cooling and Noise for Local-AI GPUs: Running 300W+ Cards Stably

2026年7月5日

I wanted to know exactly what a GPU that draws more than 300W of TDP needs — in terms of power, cooling, and noise — to run around the clock for local AI, so I looked into it. I run a dual-card setup myself, an RTX 3090 (24GB) and an RTX 3060 12GB, driving Ollama and ComfyUI on Linux every day, and these three topics are impossible to avoid.

In this article I’ve pulled together the gear and the costs needed to run a 300W+ GPU (an RTX 3090, RTX 5080, RTX 5090, and the like) stably, drawing on manufacturer spec sheets, various reviews, and user reports.

The short version: for about ¥30,000 in extra investment, you can put together a setup that runs a 300W+ GPU stably and quietly.

※ All prices in this article are rough figures based on research as of April 2026.

TDP list for current GPUs: how many watts is yours?

Here’s a rundown of the TDP (rated power draw) of the main GPUs, focused on the ones most often used for local AI.

At the 160–170W class (RTX 3060, RTX 4060 Ti), your existing power supply can often drive the card as-is, but once you go past 300W you need to rethink the power and cooling design of the whole PC. The RTX 5090 goes all the way to 575W — that’s approaching the output of a microwave oven (500–600W).

The relationship between GPU TDP and recommended PSU capacity

Just looking at a GPU’s TDP, it’s hard to tell what size power supply you should use. Below is the source data for a scatter plot. These are recommended PSU capacities assuming the whole system — not the GPU alone, but including the CPU, motherboard, storage, and so on.

[Scatter plot data] GPU TDP vs recommended PSU capacity

GPU               TDP(W)  Rec. PSU(W)  PSU price range
RTX 3060 12GB       170      550       ¥8,000–12,000
RTX 3080            320      750       ¥15,000–20,000
RTX 3090            350      750       ¥15,000–20,000
RTX 4060 Ti         165      550       ¥8,000–12,000
RTX 4090            450      850       ¥18,000–25,000
RTX 5070 Ti         300      700       ¥13,000–18,000
RTX 5080            360      800       ¥16,000–22,000
RTX 5090            575     1000       ¥25,000–40,000
RX 7900 XTX         355      800       ¥16,000–22,000

※ Recommended PSU capacity assumes one GPU + a standard CPU build (80 PLUS Gold or better)
※ For a two-card setup, simply add the second GPU's TDP on top

The key is this: against “GPU TDP + 200–250W (CPU and everything else)," choose a PSU whose capacity is 1.5–1.8× that number. A power supply peaks in conversion efficiency at around 50–80% load, so if you run it right at the edge, efficiency drops and both heat and your electricity bill go up.

[kimono_bar title="TDP comparison of major GPUs" unit="W" color="#1976d2″ highlight="5″ note="Highlight = GPU used by the author. Source: each GPU’s official spec sheet."]
RTX 4060 Ti|160
RTX 3060 12GB|170
RTX 5070 Ti|300
RTX 3080|320
RTX 3090|350
RX 7900 XTX|355
RTX 5080|360
RTX 4090|450
RTX 5090|575
[/kimono_bar]

How to choose a power supply

Pick 80 PLUS “Gold" or better

The 80 PLUS certification that indicates a PSU’s conversion efficiency comes in ranks: Standard / Bronze / Silver / Gold / Platinum / Titanium.

[kimono_bar title="Conversion efficiency by 80 PLUS grade (at 50% load)" unit="%" color="#4caf50″ max="100″ highlight="4″ note="Gold or better recommended. Source: official 80 PLUS certification standards."]
Standard|80
Bronze|85
Silver|88
Gold|90
Platinum|92
Titanium|94
[/kimono_bar]

If you’re running close to 24 hours a day, choose Gold or better. The reason is simple: a low-efficiency PSU dumps the wasted energy out as heat. That doesn’t just raise your electricity bill — it also drives up the temperature inside the case.

Specific product examples and price ranges

[kimono_product id="15848″]

A lot of people agonize over whether 750W is enough or whether to go 850W, but if you’re running a single 300W+ GPU, picking 850W is the safe call. The Corsair RM750x (about ¥15,000) is enough on paper, but you may end up buying a new PSU when you upgrade the GPU down the road. The Seasonic FOCUS GX-850 (about ¥20,000) also has a good reputation, and from what I found, either one is a safe choice.

If you’re using an RTX 5090, 1000W or more is mandatory. For a two-card setup you need the 1200W–1600W class, but by that point the options are limited and prices jump sharply.

Swapping case fans: the biggest win for quieting things down

Run a 300W+ GPU for a long stretch and the GPU temperature often climbs into the high 80s °C, at which point the GPU fans spin up to full and generate a fair amount of noise.

The first thing to do is not “lower the GPU fan speed" but improve the case airflow itself so cold air reaches the GPU. If the case fans are doing enough work, the GPU fans can hold temperature even at lower RPM.

Recommended quiet fans

[kimono_product id="15849″]

The Noctua NF-A12x25 PWM is the go-to quiet fan, said to have an excellent balance of quietness and airflow. At about ¥3,500 each, six of them come to about ¥21,000. Honestly, that’s a decent chunk of money.

The be quiet! Silent Wings 4 (about ¥3,000) is also highly rated for quietness, but a common comparison point was that the cost gap versus the Arctic P12’s pack pricing is large.

How many do you need?

For a mid-tower case, the minimum layout looks like this.

• Front (intake): 2–3 fans
• Rear (exhaust): 1 fan
• Top (exhaust): 1–2 fans

That’s 4–6 total. When you’re using a 300W+ GPU, the top priority is securing plenty of intake from the front. There are user reports that a six-fan layout — three front + one rear + two top — dropped GPU temperatures by 5–8°C.

Physical measures for the GPU

Anti-sag brackets

Large GPUs are heavy: the RTX 3090 is around 2kg, and the RTX 4090 and RTX 5090 include even heavier models. Use one for a long time and its own weight causes it to droop toward the PCIe slot — “GPU sag." That’s not just a cosmetic issue; it can also lead to poor contact and changes in airflow.

Adding a PCI-slot fan

Case fans alone sometimes don’t move enough air right below the GPU. Adding a small fan that mounts in a PCI slot improves heat exhaust off the back of the GPU.

The price is around ¥1,500–3,000. It’s not a dramatic effect, but there are user reports of an additional 2–3°C improvement.

Airflow basics

Something surprisingly easy to overlook is cable management. Especially when the power cables are a tangled mess around the GPU, they block the airflow. Just bundling the cables with zip ties (the ¥100-shop kind is fine) makes a difference.

Noise measures

Tuning the fan curve

On Linux you can set the GPU’s fan curve manually. For NVIDIA, you can control fan RPM in response to temperature with nvidia-settings or third-party tools.

Here’s a settings example that reviews and how-to articles commonly cite.

With the case fans improved and this curve in place, many reports say that during normal Ollama use (GPU at 60–70°C) fan speed stays in the 50–70% range, keeping the noise to a level that doesn’t get in the way of conversation.

Anti-vibration pads

If the case is metal, fan vibration can transfer into it and resonate. Slipping an anti-vibration pad or rubber washer between the fan and the case is said to kill that resonance.

Noctua fans come with anti-vibration pads. The Arctic P12 doesn’t include them, so you need to supply separate anti-vibration rubber pads (about ¥500 for a 4-pack). Two packs (about ¥1,000) are enough for six fans.

Using a USB fan to help with exhaust

Placing a USB fan near the case’s rear exhaust vent to push the hot air out of the room (toward a window or door) is a low-key but reportedly effective trick. Something in the ¥1,000–2,000 range is plenty.

In summer, some say whether you have this or not changes room temperature by 1–2°C. It’s more about managing the room’s temperature than cooling the PC itself, but the net result is a lower intake temperature inside the PC.

The reality of electricity costs

Measure it with a watt meter

It’s easy to say “it’s a 350W GPU, so the monthly bill is ¥X," but actual power draw varies enormously with load. The surest approach is to measure it with a watt meter.

Since it just goes between the outlet and the power cable, setup takes 30 seconds.

Power draw for an RTX 3090 + RTX 3060 build — from user measurements

Pulling together user measurements for a two-card RTX 3090 + RTX 3060 build, the numbers generally seem to fall in the following ranges.

Monthly electricity cost by GPU TDP

Here’s a comparison of monthly electricity cost for GPUs of differing TDP, running “8 hours a day of AI inference + 16 hours of standby." The rate used is ¥30/kWh.

※ Formula: (inference W × 8h + standby W × 16h) × 30 days × ¥30/kWh ÷ 1000

Comparison with cloud AI

Local AI’s electricity cost lands about the same as, or a little above, the monthly fee for ChatGPT Plus or Claude Pro. But local AI has real advantages: no token limits, fully protected privacy, and free customization. Judging it “expensive" on electricity cost alone is premature.

Summary: what you need and roughly what it costs

For running an RTX 3090-class card around the clock, I’ve organized the extras you’ll need into an example build, based on my research. If you use any 300W+ GPU, expect roughly the same outlay.

Compared with the price of the GPU itself, that’s about ¥30,000 in extra investment. Many reports say that assembling these items greatly improves quietness and makes the machine run stably even near 24-hour operation.

A high-TDP GPU won’t reliably deliver its intended performance if you just buy it and slot it in. Getting the three things — power, cooling, and noise — in order is part of the “initial investment" for a local AI setup. Plan your budget to include not just the GPU itself but this ¥30,000 as well.