RC extraction, static timing, IR/EM, thermal, stress. Five analyses, five tools, three vendors, five schedules. We run them like separate problems because that's how the tooling grew up. The physics underneath never agreed to that split.
Thermal moves both timing and IR/EM. The demand currents that drive IR come out of timing. Timing, in turn, drifts with IR drop. Pull on any one thread and the whole thing moves. So really it's one coupled multi-physics system that we've chosen, for historical reasons, to analyze as if it were five loosely related ones.
That choice costs more than convenience. A siloed tool has to guardband against everything it can't see. Timing margins pad for IR and thermal effects they can't model directly. IR analysis leans on activity profiles that may not match the real thermal-timing interaction. You end up pessimistic in some corners and, more worrying, optimistic in others, because a tool that can't see a coupled effect can't warn you about it either.
The answer here isn't another point tool. It's the layer underneath all of them. What you'd want is a backbone that runs extraction, timing, IR/EM, thermal, and stress against a single source of truth, scaling from block to full 3DIC without hopping tools or re-deriving data at every step, and doing it with real parallelism across CPU and GPU farms, because multi-physics coupling at full-chip scale isn't something you fix by making each domain faster on its own.
This doesn't mean torching decades of legacy signoff. Those solvers are strong, correlated to silicon over many years, tuned hard for accuracy in their own domain. You don't want to replace them so much as connect them: move data in and out of them fast, keep them in sync, and let each contribute its strength to a coupled result instead of working alone.
There's a quieter problem sitting behind all this. Even if you build the backbone, users still can't see what it produces, because the default output of this industry is the text report: rigid, single-domain, made for one engineer to read one number at a time. That format was never built to show a coupled result, and it shows. You can't spot a thermal-timing-IR interaction in a table designed to report timing slack. Seeing results at this scale takes the same parallelism as generating them: fast rendering, fast analytics, fast summarization across the whole dataset, not a report generator bolted on as an afterthought.
This is where a GUI actually earns its keep. An easy-to-use one that can pull up every signoff result together, across domains, in seconds rather than the hours a classic viewer needs just to open a block, turns root-causing from a multi-day hunt across five tools into something you can do in one sitting. Once you can see the timing violation, the IR hotspot, and the thermal gradient overlaid on the same layout at once, the fix usually suggests itself; finding it is mostly a matter of seeing the coupling in the first place.
The tools we have were built for a world where each domain closed on its own and got stitched together with margin. That world is mostly gone. What replaces it has to treat signoff as the coupled problem it always was, and give engineers a way to actually see it that way. That's the shift we're betting on.