Home / Learn / Data & AI / Offline AI / Lesson 6.4
All lessons

Module 6: Power, heat and running costs · Lesson 6.4 · about 15 min

Laptop batteries and thermal throttling

Why laptops and phones slow down during long AI jobs, how much battery local AI uses, and settings that keep speed steady.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 6 · about 15 minutes · read first: 6.1 and 6.3. You’ll be able to keep a laptop fast through a long AI job and know what it costs in battery.

A laptop can only go as fast, for as long, as its small cooler can carry the heat away, so long jobs need power, a performance mode and air.

1. Why this matters

Laptops and phones are where many people run their first local model. They are small, with little room for cooling and a battery to protect. Long jobs can start fast and then slow down by a third or more. Simple settings fix most of it. Think of a sprinter running a marathon: fast for the first minutes, then they must slow down to keep going.

2. Throttling

Chips have a safe temperature limit. When a long job heats them up, they lower their clock speed (how many steps the chip takes per second). This is thermal throttling. Since writing is mostly memory-bound (Lesson 4.2), the drop is often smaller than in games, but reading long prompts (compute-bound) slows a lot.

Signs: speed falls after a few minutes, fans get loud, the laptop’s underside gets hot.

A cool start. On battery, on a sofa, a 40-minute summary job starts at 12 tokens per second. The chip is cool.
Heat builds up. The soft sofa blocks the air vents. After about 5 minutes the chip reaches its safe limit.
The chip slows itself. To stay safe, it lowers its clock speed. Writing falls to about 7 tokens per second for the rest of the job.
Same job, better setup. Plugged in, on the best performance mode, on a desk with the back raised. The cooler keeps up, and speed holds at about 11.
Count the words. Tokens written = speed × time, the area under each line. On the sofa about 18,700 tokens; on the desk about 26,800, over 40% more for the same 40 minutes.

“If my laptop slows down during a long job, something is broken.” Throttling is a safety feature, working as designed. The chip protects itself and the battery by slowing down when it gets hot. Better airflow, a performance mode and mains power usually bring the speed back.

3. Battery

On battery, laptops usually switch to power-saving modes that cap the chip. A thin laptop writing with a local model may use 20 to 40 W: on a 60 Wh battery that is about 1.5 to 3 hours of steady AI use. A watt-hour (Wh) is the battery’s store of energy: a 60 Wh battery can give 60 W for one hour, or 30 W for two. Phones (Lesson 4.6) drain faster in proportion and heat up quicker.

battery hours≈battery size (Wh)power drawn (W)
Battery size (Wh)
How much energy the battery holds, printed on the label or shown in system information.
Power drawn (W)
How fast the laptop uses energy while the model writes, often 20 to 40 W.
Thermal throttling
The chip lowering its own speed to stay below a safe temperature.
Physical meaning
A battery is a fuel tank and the watts are how hard you press the pedal. The heat the chip makes is like a runner's body heat: if it cannot escape fast enough, the only way to keep going is to slow down.

4. Settings that help

  1. Plug in for long jobs.
  2. Performance mode: Windows Settings > Power > Power mode > Best performance; on a Mac laptop with the option, High Power Mode while plugged in.
  3. Airflow: use a hard, flat surface; raise the back a little; a laptop stand or cooling pad helps.
  4. Smaller load: a smaller model or quant (Lesson 8.2) and shorter prompts heat the chip less.
  5. Split long jobs into chunks with short pauses, or run them overnight on a desktop.
Maths and principle

Principle: sustained speed is limited by how much heat the cooler can remove, not by the chip’s peak speed.

Equation: battery hours ≈ battery Wh ÷ watts drawn.

Analogy: a sprinter in a marathon: fast for the first minutes, then they must slow down to keep going.

Try it: a 70 Wh battery and 28 W while writing? (70 ÷ 28 = 2.5 hours.)

5. Worked example

Made-up numbers. Ana runs a 40-minute summary job on her laptop on battery on a sofa. It starts at 12 tokens per second and falls to 7. Plugged in, in performance mode, on a desk with the back raised, it holds about 11 tokens per second throughout.

6. Practice

  1. Find your laptop’s power mode setting.
  2. Run a model for 10 minutes and note tokens per second at the start and end (Lesson 3.10).
  3. How many hours of AI use would your battery give at 30 W?
Answers
  1. Windows: Settings > System > Power & battery. Mac: System Settings > Battery.
  2. A drop of more than 20% means throttling: try the settings above.
  3. Battery Wh (on the label or in system info) ÷ 30.

New words: thermal throttling, clock speed, watt-hour (Wh), performance mode.

  • When a chip gets too hot it lowers its clock speed (thermal throttling), so tokens per second fall.
  • Writing slows less than prompt reading, because writing waits on memory rather than on the chip.
  • Battery hours ≈ battery Wh ÷ watts drawn; a thin laptop gets about 1.5 to 3 hours of steady AI use.
  • For long jobs: plug in, choose the best performance mode, use a hard surface and keep the vents clear.

Check yourself

1What is thermal throttling?
2Why does a laptop often run local AI slower on battery?
3You run a 30-minute batch job on a laptop. Which helps keep speed steady?

Sources

  • Apple, About Low Power Mode and High Power Mode on Mac, Apple Support (official page), https://support.apple.com
  • Microsoft, Change the power mode for your Windows PC, Microsoft Support (official page), https://support.microsoft.com
  • Williams, S., Waterman, A. and Patterson, D. (2009), Roofline: an insightful visual performance model for multicore architectures, Communications of the ACM 52(4), 65-76 (why memory-bound work slows less when the chip clocks down) https://doi.org/10.1145/1498765.1498785

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Prices and model names change fast: check the official source before you buy.