Keeping a 24/7 Local AI Box Quiet: Cooling, Noise, and the 275W Trick

Bottom Line

• Cap the power limit before you buy anything. On a measured RTX 3090 sweep, inference ran at 32.0 tok/s at 350W and 32.7 tok/s at 275W — the same speed. Less power, less heat, lower fan speed, zero cost.
• There is a cliff, and it is sharp. The same sweep fell to 20.6 tok/s at 200W as the core clock collapsed from 1395 MHz to 480 MHz. The useful window on that card is roughly 250-300W. Do not go lower on faith.
• Cooler design matters more than the card. RTX 3090 Founders Edition blower: 48.9 dB under load. Open-air MSI Suprim X: 31 dBA. Same silicon.
• Passive server cards are not a quiet option. A Tesla P40 has no fan at all. It expects a rack chassis. In a tower it needs a shroud and a loud fan.
• A Mac mini M4 is in a different category. Measured 24.1 dB idle and at average load, 45.1 dB only at maximum. It is inaudible in practice, and capped at 16GB or 24GB of memory.
• An honest note about the big air coolers. The Noctua NH-D15 G2 measured 44.2-44.6 dBA at maximum in review testing, and is louder at full speed than the original NH-D15. Its value is headroom, not a low peak: at inference wattage it never reaches full speed.
• If you are building multi-GPU, plan the room, not the fans. No cooling choice makes a 1,500W build quiet.

The Problem Is Duty Cycle, Not Peak Heat

A gaming rig is loud for two hours and then stops. A local AI box that hosts an always-on agent is loud continuously, and the room heats up until it reaches a new equilibrium. Those are different engineering problems, and most build guides solve the first one.

Three consequences follow:

• Average draw matters, peak draw does not. Cooling a 350W peak that occurs for eight seconds is easy. Cooling 300W sustained for eight hours is what actually sets the fan speed.
• Room temperature drifts up. Ambient rises, so the fan curve runs hotter for the same load, so it gets louder over the session. People describe this as the machine “getting louder for no reason.”
• Small noises become large. A 35 dB hum you never notice during a game is the thing you notice at 11pm.

Step 1: The Power Limit (Free, Do This First)

Local inference is memory-bandwidth bound. The GPU cores spend most of their time waiting on VRAM, so reducing the core clock ceiling costs far less throughput than it saves in watts. That single fact is why the quietest change is also the cheapest.

A published RTX 3090 sweep running Qwen 3.6 27B measured this directly:

Read the 350W and 275W rows together. The card is not faster at stock than at 275W. The 0.7 tok/s difference runs the “wrong” way and sits inside measurement variance. You are paying 75W for nothing, and that 75W is what your fans are working against.

Then read the 250W and 200W rows. The cliff is 50W wide. Between them the core clock falls by 65%. This is why “just undervolt it hard” is bad advice: below the cliff you lose a third of your speed for a modest noise gain.

Set the limit like this:

# Linux, persistent across reboots via a systemd unit
sudo nvidia-smi -pm 1
sudo nvidia-smi -pl 275

# Check it took
nvidia-smi --query-gpu=power.limit,power.draw --format=csv

Then run your model and your prompt at each step down, and stop one step above where the number drops. The cliff position depends on the card, the model, and the quantisation. Ours is not yours. Our undervolting guide has the full procedure and the Linux tooling.

Step 2: Cooler Design Beats Fan Count

The measured spread between cooler types on the same GPU is larger than anything you will gain from case fans:

Two practical rules fall out of that table.

Open-air for one card. A triple-fan card dumps heat into the case, which is fine when there is one card and reasonable intake. It is the quietest option by a wide margin.

Blower for three or more. Open-air cards stacked together recirculate each other’s exhaust and the fan curves chase it. Blowers avoid that, and they are loud. Accept that and put the machine in another room. Our multi-GPU motherboard guide covers the spacing side of the same problem.

Passive server cards are a third category. The Tesla P40 is the cheapest 24GB card you can buy, and it has no fan. It assumes a rack chassis forcing air through a bare heatsink. In a tower it needs a 3D-printed shroud and a high-static-pressure fan, and the result is not quiet.

Step 3: Airflow, Not Fan Speed

The goal is a case that moves enough air at low RPM that the fan controller never needs to ramp. Bigger fans spinning slower move the same air with less noise than small fans spinning fast.

• Front intake should exceed rear and top exhaust. Positive pressure keeps dust out of the GPU heatsink, which is what quietly degrades cooling over a year of 24/7 operation.
• Give the GPU its own intake path. A card starved of fresh air ramps its own fans no matter how good the rest of the case is.
• Filters need cleaning on a schedule. A 24/7 machine accumulates dust roughly three times faster than an eight-hour-a-day machine. Put it in the calendar.
• The CPU barely matters here. Inference load sits on the GPU. Your CPU cooler is idling.

That last point leads to an honest caveat about the coolers people buy for these builds.

The Option Most Guides Will Not Say Out Loud

If the machine must live in your bedroom or a shared office, the quiet answer is not a quieter tower. It is a smaller machine.

Notebookcheck measured the M4 Mac mini at 24.1 dB at idle and 24.1 dB at average load, reaching 45.1 dB only at maximum. In a normal room that is inaudible. It achieves this by not producing the heat in the first place, not by cooling it better.

The cost is real and you should not gloss over it. The Mac mini M4 ships with 16GB or 24GB of unified memory, and Apple removed the 32GB and 64GB options from the base M4 in May 2026. That ceiling decides which models you can run, and you cannot raise it later. Our Mac mini vs Mac Studio comparison and the soldered memory trap both cover what that limit costs you.

The pattern we see repeatedly is people buying a dual-GPU tower, discovering it is unbearable in a home office, and ending up doing their daily work on a Mac while the tower sits idle in a basement. If that is the likely ending, buy the Mac first and skip the tower. If you genuinely need the VRAM, buy the tower and plan for it to live somewhere else from day one.

Common Mistakes

1. Buying a bigger CPU cooler to fix GPU noise. The GPU makes the heat. The CPU is idling during inference.
2. Running at stock power because “the card is rated for it.” The TDP is a ceiling, not a target. On the measured 3090 sweep, 75W of it bought nothing.
3. Undervolting past the cliff. Below roughly 250W on that card you lose a third of your throughput. Measure; do not guess.
4. Choosing blower cards for a single-GPU build. You get 48.9 dB instead of 31 dBA and gain nothing.
5. Buying a passive server card for a desk machine. The P40 is cheap for a reason. It assumes a rack.
6. Ignoring dust on a 24/7 box. Cooling degrades slowly, fan speed rises slowly, and you blame the software.
7. Solving noise instead of relocating the machine. A long DisplayPort cable and an Ethernet run costs less than any cooling upgrade and works better.

See Also

• Undervolting your GPU for a 24/7 local agent — the full power-limit procedure with commands
• Local LLM electricity cost and break-even — the same watts, priced
• What your local AI rig will be worth in two years — running cooler protects resale too
• Motherboard and CPU for a multi-GPU LLM rig — card spacing, which sets recirculation
• The electrical requirements nobody puts in an AI build guide — the other constraint on a big build
• Mac mini vs Mac Studio for local LLMs — if you take the quiet route
• The soldered memory trap in unified-memory boxes — what the quiet route costs you

Sources

• RTX 3090 power limit sweet spot for local LLM inference (Jean Brito) — the full power sweep table
• Nvidia GeForce RTX 3090 Founders Edition review: power, temperatures, fan speed and noise (Tom’s Hardware) — 48.9 dB blower figure
• MSI GeForce RTX 3090 Suprim X review: temperatures and fan noise (TechPowerUp) — 31 dBA open-air figure
• The M4 Mac mini is the best desktop for most people (Notebookcheck) — 24.1 dB idle, 45.1 dB max
• Noctua NH-D15 G2 review (Tom’s Hardware) — 44.2-44.6 dBA maximum
• Noctua NH-D15 G2 review, standard/LBC/HBC comparison (Hardware Busters) — 4.3 dBA louder than the G1 at full speed

Note on comparing noise figures: the 48.9 dB and 31 dBA measurements come from different reviewers using different microphone distances and test rooms, so the exact gap is not directly comparable. The direction and the rough magnitude are consistent across every review of the two cooler types, which is the claim this page makes.