Inference, Bottom Up

Inference, Bottom Up

How a chip turns a file of numbers into text, built up one small step at a time. Use the buttons or your arrow keys.

A model is a giant list of numbers

Chip

Tokens sharing each trip: 1

Moving the weights

Doing the math

Tokens per second

All numbers use Qwen3.8 27B at 4-bit (about 16 GB) and each chip's published peak specs. Compute for the Macs and the DGX Spark is estimated: Apple doesn't publish it, and NVIDIA quotes the Spark in 4-bit. These are ideal figures. Real software reaches roughly 50 to 80% of them.