When it comes to chip manufacturing, the first thing that comes to mind is the photolithography machine. However, there is something even more concerning than the photolithography machine—if you don’t even have the pen to draw the chip, then no matter how good the photolithography machine is, it is just a decoration.
This “pen” is EDA, which you can think of as the “full-process design studio” in the chip industry. The front end is responsible for conceptualizing the architecture of the chip, just like a designer planning the functional areas of a city; the back end must implement the blueprint, ensuring that every nanometer-level “road” can be accurately traversed.
Without EDA, chip design would revert to the primitive society—drawing buildings by hand, just thinking about it is despairing.
1. EDA: The Bottleneck at the Most Critical Point
The difficulty of chip design is equivalent to building a super metropolis on a fingernail, while ensuring that hundreds of billions of “residents” (transistors) can live and work normally.
Even more terrifying is that this city is constantly undergoing “demolition and reconstruction”—from 90 nanometers to 3 nanometers, it is like increasing the capacity of the city by hundreds of times on the same fingernail. Signal interference, heat dissipation issues, power consumption control, each is a life-threatening problem.
This is when EDA needs to step in. It acts like a super steward, using mathematical algorithms for planning, and physical models for verification, turning the designer’s wild ideas into executable construction drawings.
Because of this complexity, the global EDA market has only three major players left: Synopsys, Cadence, and Siemens EDA. They have perfected this over thirty years of accumulation, making it difficult for newcomers to catch up.
2. Why EDA is More Troublesome than Photolithography Machines
No matter how expensive the photolithography machine is, at least you can see the physical object. The terrifying aspect of EDA is that it locks away knowledge, rules, and habits.
1. Experience Cannot Be RushedEDA cannot be launched just by writing some code; it needs to grow alongside chip technology. Every time the chip is upgraded, EDA must undergo major revisions—just like every time the city expands, your city planning software must be rewritten.
What we lack is not smart people, but the twenty years of accumulated experience from making mistakes. This cannot be rushed.
2. Rules are CustomizedCompanies like TSMC and Samsung, the “real estate developers”, each have their own building standards (PDK). Want to use their production line? Sorry, you must follow my rules first.
This is like needing to be proficient in the building codes of Beijing, New York, and Tokyo simultaneously; missing a single punctuation mark is unacceptable.
3. Changing Tools? It’s Easier to Change CompaniesChip design often takes a year; switching EDA tools halfway through is equivalent to making the entire team learn to speak again. The cost is so high that it makes the boss want to cry, and the risk is so great that no one dares to try.
Therefore, once you use a certain EDA, you are basically “locked in”.
3. How Accurate is America’s Seven-Inch Punch?
In 2022, the U.S. directly banned the export of the most advanced EDA tools. How severe is this move?
It’s like having all the construction site, materials, and workers ready, only to find that the design blueprints have been taken away. Even with the best photolithography machine, you cannot produce 5nm or 3nm chips.
We can now draw blueprints for 28nm, these “ordinary residential areas”, but to build “skyscrapers”, we are still far behind. The algorithms need to be smarter, the models need to be more precise, and most importantly, we need to be more familiar with the foundries—these all require time.
4. Can AI Help Us Overtake on the Curve?
Now we finally see some hope: AI is changing the game.
Previously, layout and routing adjustments that required engineers to work tirelessly can now be automatically handled by AI. The TPU chip designed by Google using AI is even better than those made by humans. Established EDA companies are also frantically training their AI models.
For us, this may be the best opportunity. The thirty years of experience of the old giants can be learned by AI in about five to eight years. We no longer have to chase behind others but can compete on the same starting line.
This path is not easy, but at least we see the dawn.