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

Assistant Workload

Requested generation rate 18 tok/s

Weight precision

Cooling solution

Ambient temperature 24°C

Live Chip State

COOL — full clock

25°Cthrottle 45°C70°C

Avg die temperature

24.0°C

Hotspot peak

24.0°C

Clock scale

100%

Actual throughput

0.0 tok/s

Power draw

0.0 W

Grid

24×24 cells

How it works

Instead of one lumped thermal mass, the die is discretized into a 24×24 grid of thermal cells. Each cell exchanges heat with its 4 neighbours (finite-difference diffusion) and loses heat to ambient through the cooling solution — the exact same physics as a single-node RC model, just spatially resolved:

c_i · dT_i/dt = g·Σ(T_j − T_i) + q_i(t) − k_cell·(T_i − T_ambient)
q_i(t) = P(t)/N_hotspot   for cells inside the compute core, else 0
P(t) = P_idle + throughput(t) · E_token(quant)

Summed over every cell, the diffusion term cancels exactly (energy leaving one cell enters its neighbour) so the grid's average temperature obeys the identical lumped equation used by the 3D version — this is a genuine spatial refinement of the same model, not a different one. The DVFS controller reads that grid-average temperature and scales the clock exactly as before:

clock_scale(T_avg) = 1                                          T_avg ≤ T_warm
clock_scale(T_avg) = 1 − (T_avg−T_warm)/(T_max−T_warm)·(1−s_min)   T_warm < T_avg < T_max
clock_scale(T_avg) = s_min                                      T_avg ≥ T_max

throughput(t) = min(demand, max_tok_s · clock_scale(T_avg))

Heat-map — live die temperature field; the compute core (centre) runs hottest, heat spreads outward and is carried away at the edges by the chosen cooling solution.
Ripples — one ring per generated token, expanding outward from the compute core at a rate equal to real throughput, not requested demand.
Strip chart — scrolling trace of average die temperature, clock scale and throughput so you can see the heat-up → throttle → equilibrium cycle unfold over time.
Weight precision / cooling / demand — identical trade-offs as the lumped model: lower-bit quantization cuts joules per token (thermal headroom), better cooling raises k_cell, higher demand pushes more power through the same hotspot.

Die heat-map (blue→amber→red) Token ripple Temperature trace Clock scale trace Throughput trace