lithos — a mini-GPU you can program in your browser

Write a kernel in a tiny SIMT assembly, run it on a real cycle-accurate simulator (compiled to WebAssembly — the same C++ engine that passes the test suite), and watch it execute: warps issue in lockstep, memory coalesces, branches diverge, and latency hides behind other warps. Learn the theory →

examples:

threads mem latency segment words

speed

cycle —

ALU memory branch idle lane divergence warps × lanes: lit cell = lane active this issue. memory: each block is a segment of words; lines run from lanes to the word they touch — many lanes → few blocks = coalesced, fan-out = scattered. data flow: real operand values flow through the op into the destination register.