Notes from the frontier.

As EDA moves toward AI agents for physical design, the bottleneck we keep hitting isn't model capability. It's that no agent today can see the full cross-domain picture the way a strong designer does.

Engineers default to text reports because classic EDA GUIs can't keep up. Pair multimodal AI with a GUI fast enough to render anything at full fidelity in about a second, and visualization finally becomes a first-class input for agents, and humans.

Every chip generates hard-won backend knowledge that rarely survives past tape-out. Carrying the underlying tradeoffs from one design to the next, not just the raw data, is one of the harder unsolved problems in the field.

The signals deep in P&R and signoff are some of the richest data a design team produces all year, yet most get used once and discarded. What if backend results fed floorplanning and RTL, instead of dying at tape-out?

We analyze signoff as five separate problems: RC extraction, timing, IR/EM, thermal, stress. The physics underneath is one coupled system, and pretending otherwise quietly costs us margin in both directions.

What if you could get a credible read on a new RTL design's congestion, timing, IR, and thermal profile before it ever goes through P&R? A fast, vendor-independent data layer is what makes cross-design training for shift-left estimation possible.

3DIC is sold as the next lever for performance and power. It's also a multiplier for every structural problem 2D physical design already has, with cross-die coupling stacked on top. Reasoning about a 3D stack means seeing the whole stack at once.