What you will be able to do

  • Explain what each major part of a computer does and how much it matters for local AI.
  • Find the graphics card, its memory, and the system memory in your own machine.
  • Tell whether an existing machine is worth trying before you buy anything.

The ranking

Lesson 9.2 gave the two numbers that matter: memory capacity and memory bandwidth. Every part below is ranked by how much it moves those numbers, or blocks you from using them.

Every part of a computer, by how much it matters for local AI
  1. Graphics card and its VRAMMost important. Capacity decides what fits; bandwidth decides speed.
  2. System RAMHolds whatever does not fit on the card. Capacity matters for big MoE models.
  3. Power supplyMust cover the card’s rated system power. Undersizing causes crashes.
  4. MotherboardNeeds a full-length PCIe x16 slot and enough RAM slots.
  5. Processor (CPU)Matters for prompt reading on CPU and for offloaded layers; rarely the bottleneck with a good card.
  6. Storage, cooling, case, OSFast SSD loads models quickly; cooling and airflow keep speed steady; OS decides which software routes are available.

Spend the budget from the top of this list down.

The parts, one at a time

Graphics card (GPU) and its VRAM. The single most important part. The card's memory holds the model; its bandwidth sets writing speed. Current cards range from 448 GB/s (RTX 5060 Ti) to 1,792 GB/s (RTX 5090). [source] A card with more memory runs bigger models; a card with more bandwidth runs them faster.

System memory (RAM). Holds the operating system, your apps, and any part of a model that does not fit on the card. Capacity matters most when you plan to run large mixture-of-experts models with the experts in RAM; 64 to 128 GB is reasonable for that. Speed comes from the memory type (DDR5 today) and from running two sticks in matching slots, so both channels are used.

Processor (CPU). Reads prompts when there is no GPU, runs offloaded layers, and runs everything else. With a capable card it is rarely the bottleneck; a mid-range desktop chip is enough.

Motherboard. Connects everything. For local AI it needs a full-length PCIe x16 slot for the card (two, well spaced, if you plan two cards), enough RAM slots, and at least one M.2 NVMe slot.

Power supply (PSU). Feeds the card. Undersize it and the machine crashes under load. Use the card maker's figure: NVIDIA lists 1,000 W required system power for the RTX 5090, whose own draw is 575 W, and supplies either four 8-pin cables via an adapter or one 600 W PCIe Gen 5 cable. [source]

Storage (SSD). Models are large files (5 to 65 GB in this path's examples). A fast NVMe drive loads them in seconds; a slow drive only delays loading, not answering.

Cooling and case. A CPU cooler (with thermal paste between it and the chip) and a case with good airflow keep the parts from slowing themselves down when hot. Long document runs keep a card busy for minutes; airflow matters.

Operating system. Decides which software routes are open. Windows and macOS have point-and-click apps; Linux is required for some server engines such as vLLM.

Apple and other unified-memory machines

On Apple Silicon Macs and some newer PCs, the processor and graphics share one pool of unified memory. There is no separate VRAM: most of the pool can hold the model. That is why a Mac mini with the M5 Pro chip (307 GB/s) can run models a 16 GB card cannot, though more slowly than a high-end card. [source] Lesson 9.5 compares these options side by side.

What is in your machine now

On Windows, open Task Manager (Ctrl+Shift+Esc), choose Performance, and select the GPU entry: it shows the card's name and its dedicated GPU memory. The Memory entry shows installed RAM. On a Mac, open the Apple menu and choose About This Mac: it shows the chip and the memory.

Then compare with LM Studio's stated minimum: 16 GB of RAM and at least 4 GB of dedicated VRAM on Windows, or Apple Silicon with 16 GB on a Mac. [source] Above that line, try what you have first (lesson 9.8).

Self-check

  1. Which part matters most for local AI, and which two numbers on it should you look up?
  2. Why does a Mac with 64 GB of unified memory run models that a 16 GB graphics card cannot?
  3. Where do you find the power supply requirement for a graphics card?
  4. On Windows, where do you see your graphics card's dedicated memory?
Check your understanding
  1. The graphics card: its memory capacity and its memory bandwidth.
  2. Unified memory lets the graphics side use most of the shared pool, so the whole model fits; a 16 GB card holds only 16 GB.
  3. On the card maker's specification page ("required system power").
  4. Task Manager, Performance tab, GPU entry.