ARCHGEN / THERMAL LAB
CONNECTING…

Same power. Same physics.
Two paths to temperature.

Compare Reference, Model 1 100M, and Model 2 27M on the same input.

Choose a map and run a live comparison. No results are pre-filled.
MethodTimeMean errorWorst cell× vs fresh Reference*
Run a comparison to populate measured results.
01 / Reference

Reference

— ms

Fresh process, grid assembly, factorization and solve.

Live reference map
appears after a run
32 × 32 surface · 3 thermal layers
02 / Final correction

Model 1 100M

— ms

The Model 2 27M prediction seeds AMG-preconditioned CG. Residual and heat balance are checked.

Corrected temperature map
appears after a run
Numerical checks will appear after a run
03 / Prediction

Model 2 27M

— ms

Coarse physics → frozen graph network → temperature preview. No retraining.

Frozen model prediction
appears after a run
Error will be measured against Reference
Shared temperature scale°C
Input / watts per surface tile

Applied power map

The same power map is supplied to both paths.

Difference / Model 2 27M minus reference

Where does Model 2 27M differ?

Blue is cooler than Reference; red is warmer. Zero is white.

Timing scope: Model 2 27M and Model 1 100M use preloaded weights and an assembled operator. Fresh Reference includes process startup and assembly. These are different workloads, so the displayed ratio is not a like-for-like solver speedup. Maps use the same corrected ambient forcing. These three synthetic demo geometries do not establish accuracy on arbitrary unseen chips. Agreement with Reference is not silicon validation.