GearClue Turns Local AI Hardware Into Arithmetic

Why it surfaced

GearClue calculates VRAM needs, electricity costs, and power-supply capacity while accounting for details such as quantization, context length, GPU transient spikes, multiple GPUs, and Korean progressive billing. Its accompanying measurements are unusually concrete: moving four of 48 layers to the CPU reportedly reduced generation speed to 49%, while a two-RTX 3060 test found actual VRAM growth 2.3% above the theoretical KV-cache figure.

A Korean site of calculators, comparison tables, and measured notes for choosing hardware for local LLMs and development.