Each explainer follows one concrete thing, such as a single prompt, through the machine. You change the input and watch the time, bytes and memory change, with numbers computed from a real run rather than drawn for illustration.
Follow one prompt through a real model running in your browser: tokens, prefill, the KV cache, decode and sampling, with the time, bytes and memory each step costs.
Free to explore; the production-scale stage is Premium.
Batching, continuous batching and the paged KV cache, run on a scheduler you can watch pass by pass, with every pass time measured on a real model.
Two stages free; the scheduler simulations and production scale are Premium.
The all-reduce behind every data-parallel training step, run on real numbers in your browser: the ring, its cost, trees for small messages, hierarchical topology and overlap with the backward pass.
Two stages free; the ring simulations, cost, topology, overlap and production scale are Premium.
Every value comes from running the system being explained. Recorded runs say where and when they were measured.
Pages open on a recorded run. Running a model on your own device is a separate choice, with the download size stated first.
Each stage links to the concept that explains it in depth. New to the field? Start with the basics.
