Junction to ambient, in one model.

Seven control volumes, one continuous solve.

Each level owns its own boundary conditions and the geometry of what sits inside it. Change a channel width and the effect travels all the way out to plant power, then comes back as a junction temperature.

Pick the rejection path, then let the weather decide the duty.

The plant is where most of the argument about a cooling design actually happens, so it is modeled as a choice rather than a constant. Each option carries its own approach behaviour, its own parasitic power and its own water draw, and every one of them moves with wet bulb and dry bulb through the year.

Economizer duty, by climate.

Free cooling is not a yes or no answer. For a given supply temperature, a site spends the year moving between full economizer, partial economizer with trim, and full mechanical cooling, and the split is what actually sets annual energy and water. Sinkpath runs the design point and then runs the year, and reports both.

The same site and the same supply temperature, run twice with only the rejection path changed, shows how much the choice of sink is worth. Austin, Texas is a useful test because it is mild enough that dry rejection looks viable on an annual average and hot enough that the average hides the problem.

Raising facility water supply by a few degrees usually moves a site between these bands more than any component swap does. That trade is the whole point of running the annual split next to the design point instead of after it.

And the case with no ambient at all.

Every option above ends by handing heat to air or to water. In orbit there is neither, and radiation to deep space is the only way out, so the terms that set the design change completely: radiator area and emissivity, view factors to the sun and to the earth, the sink temperature the surface actually sees, and the eclipse cycle of the orbit itself. Sinkpath treats this as a rejection stage like any other, which means an orbital datacenter is solved on the same junction to ambient cascade as a terrestrial one and the two can be compared on the same metrics rather than by analogy.

The efficiency numbers, computed rather than quoted.

Each of these comes out of the same solve, so they move together when you change something. PUE on its own hides too much, which is why TUE leads here.

Real hardware, not placeholder properties.

A model is only as good as what you can put into it, so these ship with the tool.

Pressure is modeled, not assumed.

Most thermal tools stop at temperature and let pump power in through a fixed allowance. Sinkpath derives the circuit from architecture and technology, then solves it: one cold plate and one connector pair per chip section at that section's own flow, pipes and manifolds sized to a design velocity, fixed bore components scaling as flow squared.

Six pressure drop model classes are available per component, from a fixed value to a fitted polynomial, a vendor table or Darcy friction. Components you have not characterized report as unmodeled instead of quietly contributing zero.

Temperature is a lifetime, not just a limit.

A design that passes at the junction temperature limit and a design that sits ten degrees below it are not equivalent, and a pass or fail verdict hides the difference. Sinkpath reports the thermal margin at every node and carries it through to device life, so a cooling choice can be argued on expected reliability rather than on headroom alone.

The same treatment applies to the plant. Approach temperatures, economizer hours and compressor lift all move with the weather, so the annual mode split is reported alongside the design point rather than in place of it.

One algorithm, checked branch by branch.

The same generic solver runs every case, with no per paper calibration and no fitted constants. Each benchmark states which values were fed and which were predicted, shows the substituted arithmetic step by step, and reports error per row with a group RMSE. Falsification cases run alongside: deliberately wrong modeling assumptions are checked to confirm they miss by margins the correct model does not.

Some branches are finished and some are still running. The table says which is which, because a validation table that only lists wins is not a validation table.

What is anchored, and what is not.

Modeling tools tend to claim every technology equally. These legs are not equally supported, so here is the honest version. An experimental anchor is the strongest claim on the list. A model anchor is weaker by construction, since a published model can be reproduced faithfully and still be wrong about the world, and it is labelled as what it is rather than folded in with the rest.

Two findings from that work changed the model itself. In single-phase immersion the fluid heat up resistance belongs at the tank boundary rather than stacked on the local heat transfer coefficient, and stacking it misses by a factor of fifteen. A case to fluid resistance quoted without its reference temperature is ambiguous by up to 39 percent at low flow, so every resistance reported here names inlet, bulk mean or outlet.

Built inside a thermofluidics research group.