thermal · conjugate heat transfer + gradient descent
Hardware
NVIDIA GH200
Status
Passes its check
Heat-sink design by gradient descent
Abstract
Gradient descent walks 20 parameters of a pin-fin heat sink downhill: each row’s pin diameter, height and offset, and the two pitches. It took 13 steps on a 0.5 mm grid; the 40 designs in the film lie along that path, each re-solved at 0.25 mm. Air enters on the left, and the particles take the temperature of the air they are in, so they warm as they pass the pins. Over the 40 designs the pins grow taller and thinner, the hottest point of the base falls from 54.6 to 38.0 K above the inlet air, and the chip’s thermal resistance falls from 10.92 to 7.60 K/W (−30%), for about 25% more pumping power.
Fermioniq·8 October 2026
Figure 1
448 × 176 × 88 cells of 0.25 mm · 40 designs in 4.3 h (~6 min each) on an NVIDIA GH200 · rendered in Blender Cycles
How it was checked
Reference
Energy balance: the heat leaving through the outlet is 4.96–4.99 W of the 5 W input in every design. Grid check: the objective falls 21% on the 0.5 mm grid and 23% when the same designs are re-solved at 0.25 mm
Caveat
Between solved designs the pins are interpolated and the colours cross-faded. The particles ride a 0.5 mm copy of each solved flow. The pins are drawn as smooth cylinders from each design’s parameters; the solver sees them on the 0.25 mm grid.
More from the reel
Simulations written with Sofia, one per physics domain, each verified against a reference.
