Package Power Density Console

The headline number on an AI accelerator is its wattage — 700W, 1,000W, 2,700W — but that number alone doesn't tell you whether you can cool it. The quantity that actually governs cooling is heat flux: watts per square centimeter. A given cooling technology can only pull heat away up to a maximum flux, and once the chip exceeds it, no amount of bigger fans or faster pumps helps locally — the heat simply can't leave fast enough.

That reframing has consequences. A 700W chip packed onto a small die can be harder to cool than a 1,000W chip spread across a large package, because the small one has higher flux. And the average flux understates the problem, because power is never uniform — compute cores concentrate heat into hotspots whose local flux can be several times the package average. Cooling must beat the hotspot, not the average, which is exactly the trap a watts-only thermal budget falls into.

Stacked against fixed physical ceilings, this is why the industry climbed a cooling ladder. Forced air tops out near 80 W/cm² — fine for decades of chips, hopeless for modern accelerators. Direct-to-chip liquid cold plates reach several hundred W/cm² and became mandatory for 700W-plus parts. Two-phase immersion cools whole boards uniformly, and embedded microfluidics — channels etched into the die or interposer — push past 1,000 W/cm² for the most extreme future designs. Each rung is more complex and expensive, taken only when the flux demands it.

This console makes the trade-off explicit: it computes your average and hotspot flux and lines them up against each cooling technology's ceiling, then estimates the resulting junction temperature. For the related problem of heat trapped inside a vertical die stack, use the 3D IC console, and to see how spreading power across a larger package changes the picture, the Package Cost console.