3DIC gets pitched as the next big lever for performance and power. What gets less airtime is that it also multiplies every structural problem we've been describing in 2D, and throws in a few brand-new ones.
For an agent, or a human team for that matter, to make good calls in a 3D stack, it has to see across dies. That's harder now than it's ever been, and for reasons that will sound familiar.
The silos multiply along Z. 2D/2.5D flows already juggle disconnected tools for placement, timing, power, and thermal. In 3DIC each die can come from a different team, a different foundry, sometimes a different company, each with its own flow. A single view across the whole stack is rare; one that also connects domains within it is rarer still.
The coupling crosses die boundaries. In 2D/2.5D, a placement change ripples into congestion, timing, and IR on the same die. In 3D, a decision on one die shifts thermal and power on the die stacked against it. Heat ignores die boundaries, and so does IR drop through the TSVs. A fix on die 2 can open a violation on die 1, and you might not find out until both are closed separately and stacked together.
The data doesn't just grow, it compounds. One die's data volume is already enormous: gates, shapes, voltage points, timing reports. Stack several and you multiply all of it, then add cross-die interfaces, TSV arrays, and bump maps on top. Handing that to any decision-maker, human or model, without a way to abstract and prioritize is a losing hand.
The room-of-engineers analogy still fits. A group where each person understands only one die, or only one domain, still can't make the best stacking call. Someone has to hold how thermal, power, and timing play off each other, not just within a die but across the stack. That's a higher bar than single-die physical design, and it's a big part of why we think agentic 3DIC is a must have.