Server Calculator · Runs in your browser

Server Calculator: can my PC run local AI?

By Tiny · how I pick

Type in your graphics card and processor, or just your memory, and get a straight answer, with an estimated speed for each model size. Nothing you type leaves your device; there's no server on the other end of this form.

Start typing your card, or leave blank for integrated graphics.
Auto-fills when we recognise your card. Edit if needed.
Two GPUs pool their VRAM, letting you run bigger models.
Auto-fills when recognised. Leave 0 for one GPU.
Apple Silicon is detected automatically (unified memory).

How this estimate works

Local models are almost always run quantized (compressed). Standard 4-bit quantization needs roughly 0.55 GB of memory per billion parameters, plus a couple of GB of overhead that grows with context length. A model runs fast on your GPU if it fits in VRAM; it runs on CPU/system RAM if it only fits there. Same result, several times slower. Apple Silicon is the exception: its unified memory lets the chip use most of system RAM as VRAM, which is why a 32 GB Mac punches far above a 32 GB PC with no discrete GPU.

The speed numbers use a memory-bandwidth model: token generation is bound mostly by how fast the hardware can read the model out of memory, so estimated tokens/sec ≈ memory bandwidth ÷ model size in memory. It's a genuinely good first-order estimate, but real speed varies with quant format, context length, and software. Treat the numbers as ballparks. If the calculator says a size fits, it fits.

Shopping for a box to run this stuff? Start with the under-$300 mini PC guide or the Proxmox picks, and be honest with yourself about whether you need local AI at all before you spend GPU money. Once you've picked a box, the power cost calculator tells you what it'll cost to keep running. See all the free tools.

Related guides