Explainers

Interactive pieces about systems that are hard to picture — some a real-time animation you can scrub and scroll, some a map you take apart by clicking it.

• 17 chapters · 7–8 minutes

  Inside the Machine

  How a prompt becomes tokens. One request through a trillion-parameter mixture-of-experts model on a rack of seventy-two accelerators: across the network, through a router that wakes sixteen experts out of eight hundred and ninety-six, and back out as streamed text.

• 5 levels · click to zoom

  Anatomy of a B300

  A top-down map of one NVIDIA accelerator that opens whatever you click. Down through the package, a compute die, a cluster, a single multiprocessor, and a stack of memory twelve dies tall — until you reach the registers, where the arithmetic actually happens.