NPU Die DVFS Thermal Throttling & LLM Token Rate Lab (3D)
Interactive 3D simulator of a phone's NPU running an on-device personal-assistant LLM: watch die temperature, DVFS clock scaling and real tokens/sec collapse under sustained generation, and see how quantization and cooling change the throttle curve.
Running a personal LLM assistant entirely on a phone's NPU — for privacy, offline access, or lower latency — means the chip's own heat becomes the bottleneck, not the network. This simulator models the die as a first-order thermal-RC system driven by the power cost of each generated token, with a DVFS controller that scales the clock (and therefore real throughput) down as the die warms past its safe threshold. Pick a requested generation rate, a weight precision (FP16/INT8/INT4) and a cooling solution, then run a sustained generation burst and watch die temperature, clock scale and actual tokens/sec play out live — exactly the trade-off that makes aggressive on-device quantization as much a thermal-management decision as a memory one.
Watch a phone's NPU running a personal-assistant LLM heat up under sustained generation: a thermal-RC model and a DVFS controller throttle the clock — and real tokens/sec — as the die crosses its safe temperature.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install