Ruggi.SiteHome AI server planner
Choose what you want the machine for and which model size you are aiming at. Get what to spend on first, the memory it needs, the power it will draw and what it costs to run, with noise and heat notes. No product prices.
1. What it is for
2. Power and running cost (from your own numbers)
The watts shown are examples to edit. Look up the rated power (TDP or board power) of the parts you consider and, better, measure with a plug-in power meter. Load hours are time the GPU is actually generating, not the time you sit in front of it. Hours under load and idle should add up to 24 or less.
Your plan
What to spend on first
Noise and heat
How it works
Model weights need about parameters × bits per weight ÷ 8 GB. The conversation memory (KV cache) is the model's cache size per token × the context length, and it is held separately for each person using it at the same time. Required GPU memory = weights + conversation memory + 0.7 GB for the program, plus anything else you load. These are the same formulas as the "Can my PC run this AI model?" tool, and the figures are rough (about ±30%). A model that does not fit in GPU memory can still run with part of it in system RAM, but slower, because system RAM is much slower than a GPU's memory. Speed ceiling = 0.6 × memory bandwidth ÷ size of the active weights, the usual memory-bound approximation. Energy per month = (load watts × load hours + idle watts × idle hours) ÷ 1000 × 30 days, × the price per kWh. Almost all electricity a computer uses ends up as heat in the room.
Model data last checked . Parameter counts and cache sizes come from public model pages and are rounded. This is a planning aid; check each part's own specifications, including power supply and case compatibility, before you buy.