Module 4: The parts that matter for AI · Lesson 4.4 · about 15 min
Storage, RAM, power supply and cooling
The supporting parts of a local AI machine. How much disk space models take, how much main memory you need, how big a power supply, and why cooling matters.
Module 4 · about 15 minutes · read first: 4.1. You’ll be able to spec the rest of the machine around the memory.
1. Why this matters
The other parts decide whether the machine is nice to use: models that load in seconds, a PC that does not shut down under load, and fans you can live with. Think of a good chef. A fast knife (the graphics card) matters, but so do a big fridge (disk), a clear counter (main memory) and a strong gas line (power).
2. Storage
- Model files are big: Gemma 4 E2B about 4.6 GB, Gemma 4 12B about 7.7 GB, Qwen3.8-27B about 18 GB, gpt-oss-120b about 65 GB.
- You will keep several. Plan 100 to 500 GB free on an NVMe SSD (a fast solid-state disk) for models.
- A fast SSD loads models in seconds. A hard disk works but takes minutes. Once loaded, the disk does not affect writing speed.
- Macs charge a lot for bigger internal storage; a fast external USB-C or Thunderbolt SSD is fine for model files.
Load time is simply size ÷ read speed:
“A faster SSD makes the model write faster.” The disk is only used to load the model into memory at the start. After that, every token is read from memory, not from disk. A faster SSD shortens the wait before the first answer, nothing more.
3. Main memory (RAM) on a PC with a graphics card
Even when the model runs on the card, it is loaded through main memory, and your other apps need room too. Rule of thumb: main memory at least equal to graphics memory, ideally double. A 16 GB card wants 32 GB of main memory.
4. Power supply
A graphics card’s rated power (board power) is roughly what it draws at full load: about 180 W for an RTX 5060 Ti, 350 W for an RTX 3090 and 575 W for an RTX 5090. Add 100 to 200 W for the rest of the PC, then headroom (spare capacity, so the supply is never at its limit). Use the card maker’s recommended power supply size and the right power cables. Module 6 turns watts into money.
Here is how to size one, for a used RTX 3090:
- Board power
- The most the graphics card is designed to draw, in watts.
- Rest of PC
- Processor, memory, disks and fans together, usually 100 to 200 W.
- × 1.3
- Headroom: at least 30% spare, for spikes and a cooler, quieter supply.
- Physical meaning
- The power supply is the pipe that feeds the whole machine. If it is too narrow, the moment the model starts writing and the card pulls hard, the PC can simply switch off.
5. Cooling
Language models load the graphics chip in bursts: hard while writing, idle while you read. Long jobs (batch summaries, embeddings) load it for hours. A case with good airflow keeps the chip from slowing itself down (Lesson 6.4) and keeps fans quieter.
Maths and principle
Principle: size the supporting parts for the model, with headroom: disk for several models, main memory at least the card’s memory, power supply above the peak.
Equation: power supply ≥ (card power + rest of PC) × 1.3 or more.
Analogy: a good chef needs a big fridge (disk), a clear counter (main memory) and a strong gas line (power), not only a fast knife.
Try it: a 350 W card in a PC whose other parts draw 150 W? ((350 + 150) × 1.3 = 650 W, so a 750 W supply is a comfortable choice.)
6. Worked example
Made-up numbers. Omar adds an RTX 5060 Ti 16 GB to an office PC with a 300 W supply and 16 GB of main memory. The card’s maker asks for 550 W. He budgets for a 650 W supply and a second 16 GB memory stick. Without the supply upgrade the PC switches off when the model starts writing.
7. Practice
- How much free SSD space do you have now?
- Your PC has 16 GB main memory and you buy a 24 GB card. What should you add?
- Does a faster SSD make tokens come out faster?
Answers
- If it is under 100 GB, plan an external SSD for models.
- At least another 16 GB of main memory (to 32 GB or more), and check the power supply.
- No. It only shortens loading time.
New words: NVMe SSD (fast solid-state disk), board power (a card’s rated draw), headroom.
- Keep 100 to 500 GB of fast SSD space for models; the disk only affects loading time, not writing speed.
- Main memory should be at least as large as the graphics card’s memory, ideally double.
- Power supply ≥ (card + rest of PC) × 1.3, rounded up to the next size, and at least the card maker’s recommendation.
- Good airflow stops the chip slowing itself down and keeps fans quieter.
Check yourself
Sources
- Model download sizes: Ollama library, read 16 November 2026 via BenchLM, https://benchlm.ai/best/ollama-models
- Graphics card board power: NVIDIA product pages via Compute Market, updated 1 November 2026, https://compute-market.com
- llama.cpp project README: model files are memory-mapped from disk when loaded (official page), https://github.com/ggml-org/llama.cpp
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.