Token TideBatch Dispatch
One accelerator. Eight memory blocks. A city waiting for its replies.
Make room for the whole wave.
Choose waiting request cards and Dispatch batch. Requests in one batch must use the same model and total at most 8 memory blocks. They all finish together. Select several compatible cards to save energy.
Clock cost = longest reply + 1 tick when loading a different model. Your first run is also a cold start. A warm model stays loaded through idle time. Energy cost = 2 + longest reply, regardless of the number of requests in that batch. Idle time costs no energy. No overlapping batches.
Every card shows its arrival, memory, reply length and finish-by deadline. Finish exactly at a deadline to serve it. Later means missed; a missed request leaves the queue. New requests can arrive while the accelerator runs. Use Wait 1 tick when needed, or to see how a deadline passes. Time advances only when you act.
Serve everyone for ★★★. At least 75% earns ★★; at least half earns ★. Undo and retry freely. A hint finds a feasible plan and selects its next batch; you still dispatch it. Saves and medals stay on this device.
A fictional scheduling puzzle inspired by the 2026 inference boom. These small integer costs are game rules, not hardware measurements. No models run and no requests leave your browser.