Fig 9.1The time trial (a design)Buttons or keys 1–44 lanes · 3 campaigns
Pick a scene, or wait and the track will play them in turn.
The dashboard
A decode speed+0.0%0 runs
B prefill speed, decode guarded+0.0%0 runs
C fit in 12 GB+0.0%0 runs
The track: four lanes, one P100 each. The gate and the timers are the harness the agents can't touchBlue, green and violet: three campaigns, each with its own goal and its best car in the caseThe board: one dashboard over every campaign, shown up close beside the drawingA lit car: a confirmation, a promising experiment run again on another laneThe audit beacon: a result too good to be trueA drawing of a design, not of running hardware: the research harness isn't built yet. In the plan each lane is one of the GPU host's P100s (the number of lanes is set in the cluster's manifests), each car is one experiment, a heat stands for a round of short screens that take minutes each, the case holds each campaign's confirmed best, the board is the planned Grafana dashboard over the results ledger, and the slips are its receipts. In the drawing every campaign races for time; in the design each campaign's contract names its own goal, guards and limits. Run times come from a small physical model of this track.