Hardware
The machines underneath the measurements, and what changes when you swap one part of them.
In this section
GPU
What a given card can and cannot run. VRAM ceilings, driver versions that move them, and the point where more memory stops helping.
Apple Silicon
Unified memory against dedicated VRAM, measured with the model, the quantization and the context length held fixed across both machines.
Storage & RAM
The parts that decide how long a model takes to load and whether offloading is survivable. Tagged storage or ram.
Everything in this section
Nothing published here yet. This page stays out of the index until an article lands in it.