Noise and power of a mini PC left on all day

By Emmanuel Sauvage · Published · Updated

This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.

Running Ollama all the time turns a mini PC into a small family server. Open WebUI (a web page for talking to Ollama from a browser on the home network) is there in the evening, automations call a local model, and sometimes a Whisper transcription runs in the background. GGUF files, the model format Ollama loads, often stay in memory in Q4 (a light quantization, a good size-to-quality trade) so the next reply is fast. The cost is not only the case you bought. It is decibels in the living room and watts on the electric meter.

Keep these ranges (October 2026, a typical household)

  • Ollama idle (a 7B–8B model loaded, nobody asking a question): often 12–28 W on a CPU mini PC, 18–35 W with a small RTX.
  • Living-room noise: an office mini PC is often 25–35 dB(A) at idle. dB(A) is the A-weighted decibel figure on a spec sheet. A fanless box or a Mac mini M4 is quieter. Detail is below.
  • The bill: about 20 W on average is roughly €55 a year at €0.25/kWh. The formula and the tables are in the electricity budget and the 3-year comparison.

This guide separates what a PC actually draws when Ollama is loaded and waiting from a plain system idle, and it gives rough figures so you can balance quiet, budget, and speed.

Why leaving it on all day changes the math

Occasional use hides the cost. You turn the machine on, you talk to the model for about twenty minutes, you turn it off. Left on all day, it stays powered even when nobody is asking it anything. Ollama often keeps a process running, and it may hold the last loaded GGUF in memory so the next message is faster. The draw you care about is not a Windows PC in sleep. It is a light server with a model ready to answer.

On noise, a gaming mini PC or an RTX tower spins up as soon as a GPU generation starts. Left on all day, short nighttime spikes (a home-automation script, an automatic summary) can surprise you if the case is in a bedroom. Before you buy, read this next to choosing hardware, the mini PC comparison, and the selector if you are choosing between Budget and Comfort.

Ollama idle versus system idle

Mainstream comparisons often quote “5–15 W in system idle.” That measurement is the operating system running, the screen off, and no heavy background service loaded. At home, the useful reference is:

  • System idle: Ollama is stopped, or no model is loaded. Disk and RAM use are minimal.
  • Ollama idle: the service is started, a typical model (often a 7B–8B in Q4) is already loaded, and it is waiting for a request. The GPU or iGPU may stay partly awake, depending on drivers and settings.
  • Inference: tokens are being generated. Watts and fans peak.
  • A short spike: a model download, a compile, or the first request after you change tags.

To measure at home, note the power at the wall with a plug-in power meter. If you do not have one, use your utility’s usage graph, and only if nothing else in the house is drawing power. Wait five minutes after loading a model with ollama run, without asking a question. Then compare that with Ollama stopped. The gap is what you pay to keep the assistant ready all day.

Before the detailed table: use an Ollama idle row, not system idle alone, for the yearly kilowatt-hours. An RTX speeds up a reply and often raises the idle floor. The Mac mini M4 remains a quiet reference for moderate chat (Mac mini M4 for local AI).

Rough watt figures (household mini PC, October 2026)

On a phone, scroll sideways if the columns run past the screen.

These figures are not a promise. They vary with the CPU (Intel N, Core Ultra, Ryzen), whether an RTX is installed, BIOS settings, and the Ollama version. Use them to compare scenarios, not to promise an exact bill.

State Mini PC, CPU only (32 GB) Mini PC plus RTX 4060 or 4070 Mac mini M4 (32 GB)
System idle (Ollama off) 5–12 W 8–18 W (GPU asleep) 4–10 W
Ollama idle (7B–8B Q4 loaded, no request) 12–28 W 18–35 W 10–22 W
Chat inference, 7B–8B (GPU or iGPU) 35–85 W 55–140 W 15–45 W
Inference at 13B or larger, or a long context 60–120 W 90–200 W 25–70 W
Spike (model pull, heavy stress) 80–150 W 120–250 W and up 40–90 W

What matters:

  • The Ollama idle row is the one to use for a year of always-on use, with an assistant that is always ready.
  • An RTX improves speed and often raises the idle floor if the NVIDIA driver keeps the card awake.
  • The Mac mini M4 is often the quietest in a living room for moderate chat. See Mac mini M4 for local AI for the limits on large models and images.

Noise: decibels, where you put it, and the case

What you hear depends on airflow (a 40 mm fan versus a 120 mm fan), GPU load, and the furniture. A mini PC in a closed sideboard resonates and runs hot. Practical rules for a household:

  1. Give it air. Leave at least 10–15 cm (about 4–6 inches) of open space around the vents. Do not let dust pile up. Clean it once a year.
  2. Keep the model in check. An 8B Q4 runs cooler than a 13B at full load. Line this up with VRAM and model size.
  3. Plan the hours. If you mostly use it in the evening, shutting it down or using a deep sleep overnight can be worth it. See the cost section below.
  4. Fanless means no fan. For light text and a modest Ollama idle, some fanless mini PCs (N100, N305) stay silent, and they are slower when generating a reply.

Rough subjective figures, 1 meter away, in a quiet room:

Profile Noise at Ollama idle Noise during a reply Hardware lead
Fanless N100 Nearly inaudible Slight warmth, slow CPU Example search: fanless mini PC on amazon.ca (partner link — how affiliation works)
Office mini PC, 35 W TDP 25–35 dB(A) 35–45 dB(A) Mini PC comparison
Mini PC plus a compact RTX 30–38 dB(A) 40–50 dB(A) VRAM and model size
Mac mini M4 Often 20–30 dB(A) A moderate rise Mac mini M4 for local AI

TDP is the chip’s thermal design power, a rough heat budget, not a wattmeter reading. Read the idle and load noise figures on the spec sheet, and check the current price, rather than relying only on a mainstream CPU benchmark such as Cinebench, a lab score of processor load.

Kilowatt-hours per year

Formula: kWh/year ≈ average watts × 24 × 365 / 1000.

Examples that use Ollama idle as the main assumption (the machine is always ready):

Average watts (assumption) kWh/year At about €0.25/kWh (rough)
15 W ~131 kWh ~€33/year
25 W ~219 kWh ~€55/year
40 W ~350 kWh ~€88/year
60 W (RTX, mixed use) ~526 kWh ~€132/year

If you add an hour of heavy generation a day, the average goes up. If you shut the box off for 8 hours at night, count about two thirds of the idle watts in the yearly figure, because it runs 16 hours instead of 24. For rates, off-peak hours, and a comparison with a ChatGPT subscription, open the electricity budget for a home AI box. It reuses these watt figures and lines them up with the 3-year cost comparison.

Kilowatt-hours. A mini PC at 20 W on average over the year is often less than an old, poorly insulated refrigerator, and more than a home internet router. Measure a real week after you install, instead of copying a table. Set your own rate in the hardware selector before you compare with the cloud over 36 months.

Typical always-on setups

Family assistant (Open WebUI). Open WebUI is the usual web page for talking to Ollama from the living room. Ollama idle with Llama 3.1 8B, or an equivalent, and spikes when teenagers ask about homework in the evening. Aim for 32 GB of RAM, an NVMe SSD for the models (SSD storage), and a living-room spot if the noise stays acceptable.

Home automation plus a language model. Home Assistant or n8n send short requests. The machine can sit for hours at Ollama idle between two automations. A NAS can be enough for tiny models. Comfortable replies still belong on a mini PC or a Mac.

A developer using Continue. Continue is an open-source completion and code-chat extension for VS Code, pointed at the Ollama API. The GPU stays busy longer. Leaving the machine on all day pays off mainly if you keep the model warm. Otherwise, shutting it down between sessions cuts noise and the bill. Dedicated guide: local coding with Continue.

A Docker stack (Ollama, Open WebUI, and other isolated containers) adds 2–8 W at idle. See Docker and Ollama.

Cut power and noise without shutting everything off

  • A smaller default model. Use a 3B for home automation and an 8B when you want a better chat.
  • Run some layers on the CPU if VRAM is tight. It is slower, and often cooler while the GPU sits idle. Detail is in VRAM and model size and LM Studio.
  • Ollama and the operating system. Do not leave several large models loaded. Watch ollama ps.
  • A scheduled sleep. A scheduled task, or a smart plug, can cut power overnight if nobody is using the service.
  • A careful undervolt on the CPU (PCs only, not a Mac). The gains are modest, and a bad setting can make the machine unstable.

Mac mini, gaming tower, and NAS

Option Quiet Power, left on all day A comfortable language model
Mac mini M4 Excellent Low to moderate Chat, not a large FLUX job
Fanless mini PC Excellent Low Small models
Mini PC plus RTX Varies Moderate to high Faster tokens
ATX tower Often loud High The most flexibility
Synology NAS Discreet Low Small models, CPU only

Install and measure after you buy

  1. Install Ollama, or follow a first local AI in 15 minutes.
  2. Download the model you want (choose a first model).
  3. Measure Ollama idle for 10 minutes, then a 15-minute chat. Note the peak watts and the average.
  4. Move the box, or change the model, if the noise is past what the household will tolerate.
  5. Enter the estimated kilowatt-hours in the hardware selector with your price per kilowatt-hour.

Safety and privacy, in short

A machine left on all day should not be exposed on the internet without a login. Keep Ollama on the home network, or behind a VPN. Put guests on their own Wi-Fi. Running locally keeps prompts in the house, and a model can still invent facts. Do not use it in place of a professional for health or legal questions. See privacy when AI stays home if you store sensitive documents.

Conclusion

The right mini PC for all-day AI is the one whose Ollama idle is acceptable to the ear and on the bill, not the one that wins a one-minute benchmark. Measure it, compare it with the electricity budget, prefer fanless or a Mac mini if the living room is quiet, and accept an RTX only if you need the speed or local images. Rough ranges. Check the current price (October 2026) before you buy.

Common questions

Can I leave Ollama on 24/7 in the living room?

Yes, if the vents are clear and you accept the kilowatt-hour cost. Measure watts at idle with the service running, not only with the operating system sitting idle and no model loaded.

How do I cut the noise without turning the machine off?

Limit GPU power, pick smaller models, move the case out of a closed cabinet, or aim for a fanless mini PC or a Mac mini.

Is a NAS quieter than a mini PC for AI?

Often quieter under a light load, and slower. See the Synology NAS guide for the tradeoffs.

Sources

Narrow it down with the hardware selector · French version of this section