Chip Design in 2026: An Engineer on What Changed Since 2021

Archi3 min read

semiconductorsengineeringsupplychaininterview

Contents
  1. Five years in a chip design lab
  2. The starting point: 2021 versus now
  3. Tools, AI, and the new cost structure
  4. Verification and the schedule squeeze
  5. The supply chain and geopolitics

Five years in a chip design lab#

We spoke with a senior physical design engineer at a mid-sized fabless semiconductor company. They have worked on chips for consumer devices and automotive sensors since 2018. The conversation has been edited for length and clarity.

The starting point: 2021 versus now#

Q: What was the biggest difference in your daily work five years ago?

A: In 2021, we were still heavily reliant on manual iteration for floorplanning. A single block might take two weeks of back-and-forth with the timing team. Today, the same block takes four days because the place-and-route tool suggests legal placements that meet timing 80% of the time. The tooling has improved, but the license costs have roughly doubled. Our EDA budget went from $2.1 million to $4.3 million annually. That number is now a board-level topic.

Q: Did the chip shortage change how you plan?

A: Yes. In 2021, we designed for performance first and worried about substrate availability later. Now we start with a substrate and packaging availability check before we finalize the architecture. Our last project had a six-week delay because a specific flip-chip package was allocated to a larger customer. We redesigned the package and lost two months. That would not have happened five years ago.

Tools, AI, and the new cost structure#

Q: You mentioned AI assistants in design. How do they compare to 2021?

A: In 2021, we had rule-based scripts for repetitive tasks. Now we have models that propose timing constraints and flag likely DRC violations before signoff. The accuracy is around 70% for common blocks. It saves maybe 15% of engineering hours. But the compute cost is real: our AI-assisted verification runs on a cluster that consumes 40% of our monthly cloud budget. Five years ago, that budget was mostly for regression tests, not model inference.

Q: What about talent?

A: In 2021, we could hire a junior physical design engineer in three months. Now it takes six to nine months, and the salary expectations are 30% higher. The pool has not grown much. Universities are producing more graduates, but many go into software or AI infrastructure, not silicon. We have started training technicians to run tool flows, which frees up engineers for architecture work. That was rare five years ago.

Verification and the schedule squeeze#

Q: Has verification changed?

A: The volume of verification has exploded. In 2021, a typical block had 20,000 coverage points. Now we track 60,000. The tools are faster, but the simulation farms are larger. We used to finish regression in 48 hours. Now it takes 72 hours even with more machines. The bottleneck is not compute; it is the time engineers spend triaging failures. We added a dedicated triage team last year, which we did not have in 2021.

Q: What about analog and mixed-signal?

A: That is where the gap is widest. Five years ago, analog designers could rely on corner runs that finished overnight. Now, with more corners and higher frequencies, a full corner sweep can take a week. We have started using surrogate models to predict which corners are likely to fail, but adoption is slow. Most analog designers I know still trust the full sweep. It is a cultural thing.

The supply chain and geopolitics#

Q: How has the supply chain changed your design choices?

A: In 2021, we designed for a single foundry at 16nm. Now we dual-source at 12nm and 16nm, which means two sets of design rules and two verification flows. The engineering overhead is significant. We estimate 20% more engineering time for the same function. But the risk of a single-source disruption is too high. Five years ago, that redundancy was a nice-to-have. Now it is mandatory for any automotive customer.

Q: Any final comparison?

A: The work is more expensive and more constrained, but also more predictable in some ways. The tools catch more errors early. The trade-off is that you need more capital and more patience. In 2021, a small team could tape out a chip for $5 million. Now the same chip costs $8 million to $9 million when you include tool licenses, verification compute, and packaging deposits. That changes which projects get funded. It is a different industry.

More articles