NASA gave CEA a subroutine interface, and a 15-minute sweep became 1.11 seconds

A 40 percent penalty worth about 0.004 seconds implies CEA2 needed roughly 0.010 seconds for a standalone call and CEA v3 needs about 0.014 ^[7][1]. Then take the benchmark sweep: 108,500 cases in about 1.11 seconds is roughly 10 microseconds per case ^[8][2], about 1,370 times faster than the same code answering one case on its own ^[4]. The equilibrium solver did not get 1,370 times faster. The cost of asking it a question went away.

The other half of the diagnosis is in CEA2's own number. Fifteen minutes across 108,500 cases is about 8.3 milliseconds per case ^[3], which is the same order as its single-shot time ^[1]. A sweep bought no amortisation at all, because there was nothing to amortise: with no subroutine interface ^[3], a parametric study was 108,500 separate program invocations, each parsing its own input file. The stated 800-fold speedup ^[8] is not a solver result, it is the removal of that per-call overhead, and it checks out against the raw times at about 811 ^[5].

What the new architecture adds on top is aimed at the same problem. The solver is described as thread-safe, with object-oriented data structures and stricter typing ^[6], which is what lets a design-of-experiments driver fan cases across cores instead of queueing them. Analytic total derivatives ^[10] matter for a narrower reason: a gradient-based optimiser that cannot ask for sensitivities has to reconstruct them by perturbing inputs and re-solving, which multiplies the case count and imports step-size guesswork into the answer. There is also support for negative reactant amounts and for inert hydrocarbon representations of RP-1, Jet-A and the JP-series fuels ^[10], the kind of accounting that engineers previously handled outside the tool.

The legacy command-line path is the part that decides whether any of this gets adopted ^[5]. Preserving existing input decks means the migration argument is no longer about re-validating a heritage tool against decades of practice; the validated technical basis and the files that exercise it come forward, and NASA's guidance to use v3 for new analyses ^[11] costs a team four thousandths of a second per interactive case ^[7]. That is a smaller price than most organisations pay to keep a 2002-vintage procedural code buildable ^[1].

Which leaves process, not software, as the obstacle. Propulsion thermochemistry has been the step in multidisciplinary loops that ran by hand because it could only be driven by files, and the bulletin's contact point for CEA v3 is Mark K. Leader at Glenn Research Center ^[12].