Home▸AI & Machine Learning▸NPU Thermal Grid: 2D Die Heat-Map & DVFS Throttling

NPU Thermal Grid: 2D Die Heat-Map & DVFS Throttling

2D finite-difference heat-map simulator of a phone NPU die: watch a lumped thermal-RC + DVFS clock-scaling model spread across a discretized grid, with token-generation ripples and a live temperature/clock/throughput strip chart.

AI & Machine Learning2DModerate60 FPS📱 Mobile-adapted⇄ 3D version

The same on-device LLM thermal-throttling problem, seen as a 2D field instead of a single number: a 24×24 finite-difference grid discretizes the NPU die, with a compute-core hotspot at the centre injecting power and heat diffusing outward to cells that shed it to the cooling solution at the edges. The grid's average temperature drives the identical DVFS controller — clock (and real tokens/sec) scales down once that average crosses the warm threshold and floors at the thermal limit. Ripples expanding from the hotspot mark each generated token at the actual throttled rate, and a scrolling strip chart tracks temperature, clock scale and throughput together so the heat-up → throttle → equilibrium cycle is visible over time, not just as an instantaneous readout.

⚙ Under the hood

A 2D finite-difference heat-map counterpart to the 3D NPU thermal simulator: a 24x24 grid of thermal cells diffuses heat from a central compute-core hotspot to the cooling solution at the edges, while the grid's average temperature drives the same DVFS clock-scaling controller. Token generation appears as expanding ripples from the hotspot, and a scrolling strip chart tracks temperature, clock scale and throughput together.

on-device LLMNPUthermal throttlingDVFSheat diffusionquantizationedge AI

2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install