QUANTA · INFERENCE SERIES
Four theories, four experiments you can play with.
Skip the equations — the sliders, switches and animations will read you the paper.
One request, two kinds of physics.
Twin swim-lanes race a monolithic chip against a split design. Drag the prompt-length slider and watch TTFT and throughput diverge.
Enter the lab ↗Memory, layered by temperature.
A live simulation of KV-block lifecycles: into context, down to the PIM hot tier, into the cold pool — then one recall lights it back up.
Enter the lab ↗Requests land on the cheapest tier.
Fire four kinds of requests and watch where each lands; then pull the offline switch and feel degraded mode for yourself.
Enter the lab ↗One line that sees through a machine.
A living roofline chart: drag the bandwidth and compute sliders and prove that 10× compute does nothing for decode.
Enter the lab ↗The arguments behind these labs live in Chapter 6 of the Noeton paper.