Reborn in 2010, the AI ​​in my mind is too powerful.

When Chen Yiran returned to the Samsung booth, the crowd had finally thinned out a bit.

An engineer was drinking coffee next to the Exynos processor display board, while several staff members in dark blue vests were tidying up white papers inside the booth.

After Chen Yiran went inside, he randomly pulled aside a staff member, identified himself, and expressed his desire to discuss cooperation.

Unexpectedly, the other party knew FlowEDA and Chen Yiran's name. They asked him to sit down and wait a moment, and then called out the vice president of Samsung's chip foundry business who was leading the team this time.

This was a South Korean man in his forties named Park Jae-sun. He was very enthusiastic towards Chen Yiran and immediately started praising her.

"I visited your booth in the North Hall on the first day. I also had a chat with the technical team from Korea yesterday. I was planning to visit you later today, or maybe after the exhibition to discuss cooperation with your company."

"Sorry, there were too many people in front of Grey's booth the other day, so I didn't notice you. We are also very interested in cooperating with Samsung, and we hope to get the opportunity for PDK certification, preferably as soon as possible."

"Haha, you've been to TSMC, right? Don't worry, we're much faster than them. To be honest, I admit their technology is a bit better than ours, but you can't blame our foundry division entirely; we're just as good."

Park Jae-sun's complaints are not without reason; Samsung's inability to compete with TSMC in foundry services is not the fault of the foundry department.

Since its founding in 1987, TSMC has ingrained the principle of not competing with its customers into its DNA. Although chip manufacturing may seem high-end, their corporate culture is actually quite similar to that of Foxconn.

Morris Chang defined TSMC's four core values ​​as: integrity, commitment, innovation, and customer trust.

All three of them essentially boil down to one thing: to sincerely move the client, and they certainly succeeded in doing so.

While there have been instances of TSMC employees leaking production dates to profit in the stock market, leaked design blueprints have never occurred. TSMC's integrity management system is the strongest in the industry, which naturally earns the trust of its customers.

In comparison, Samsung's position is rather awkward. Samsung designs its own chips, and even has a very complete product range: Exynos processors, DRAM controllers, NAND controllers, and image signal processors.

When Qualcomm outsourced the Snapdragon design to Samsung for manufacturing, it was essentially handing over its most core technology to a competitor that also made mobile processors.

In 2010, Apple outsourced the manufacturing of its A5 and A6 chips to Samsung, but that was because only Samsung could handle such a large volume. Even Qualcomm, while also choosing Samsung for manufacturing, has always had a complex and often cutthroat relationship with the company.

If TSMC can take over the production capacity, Apple, Qualcomm, and others will immediately replace Samsung with TSMC.

In terms of corporate culture, Samsung is destined not to completely adopt a supplier mentality.

In terms of investment scale, the difference is even more stark. TSMC is entirely a foundry; it has no other avenues to explore. Samsung, on the other hand, is a behemoth, even though TSMC's market capitalization later became nearly twice that of Samsung.

But you can only look at stock market capitalization as an example. Those old money companies in Hong Kong that pay out billions in dividends every year have market capitalizations that are nothing to look at.

Samsung is a massive and formidable conglomerate in South Korea, spanning everything from banking and insurance to food and agriculture, making it nearly impossible for it to focus its attention entirely on a single area.

Even within the semiconductor sector, departmental investments are prioritized for memory chips. The actual investment in foundry services is far less than that of TSMC.

All these influences have been applied to EDA (Electronic Design Automation), EDA tools, IP libraries, design service companies—it's fair to say that the entire semiconductor design ecosystem implicitly prioritizes TSMC. All new process certifications will be awarded to TSMC first, with Samsung following behind.

Chen Yiran felt that Samsung's miserable situation seemed more suitable for him; the industry leader was already complacent and no longer valued him.

Then they should go all out against the second-largest player in the dairy industry, muddying the waters so that they can raise prices.

"TSMC is asking for 9 to 15 months. Can Samsung deliver faster?"

Park Jae-sun was very sincere, so sincere that Chen Yiran found it a bit unbelievable.

"TSMC's process for new EDA tools is just like that; it's not targeted at you. That's because they're the industry leader. All the big EDA companies are eager to work with them, and they don't care about new, smaller companies. We're different. Although we also have framework agreements with Synaptics and Cadbury, we're happy to try new tools."

I can promise you that we will complete PDK certification as soon as possible, but we hope to sign an agreement that, for the next five years, all your technological updates must be completed before TSMC's.

Chen Yiran thought for a moment and said, "I have already given GlobalFoundries a guarantee. Although we haven't signed a contract, I am ensuring that they will be in the first tier of technology and innovation. So the speed of your two companies is definitely about the same."

However, I can confirm that we are unlikely to have any further contact with TSMC in the near future. After all, this contact wasn't very pleasant.

"Oh, that's alright. As long as it's faster than TSMC, that's fine. And to be honest, this is something that will only come up if it's truly proven to be very successful. There's another requirement, or rather, our cooperation request, which is the more important one."

Chen Yiran was a little confused. He didn't know what else Samsung could ask him to cooperate on besides EDA.

Park Jae-sun sighed and said, "Sigh, we're still not close enough to Silicon Valley, and we're not quick enough to react to many things, but at least we react faster than TSMC, and that should be enough."

Park picked up a printed document from the table. It was an arXiv preprint of Andrew Ng's paper, with the parts of the Methods section that cited DeepFlow highlighted in a highlighter.

"Samsung's team has carefully studied the data from the Stanford lab, but they've only been studying it for a week and haven't come up with complete results yet. However, after seeing your grey at the exhibition that day, and especially after confirming some routes with the engineers in Korea, we have a preliminary conclusion."

Chen Yiran still didn't understand what Samsung meant.

Park Jae-sun said with a smile, "We're running our own OPC script on the NVIDIA Fermi architecture, but we're stuck on PCIe data transfer, and the speedup is only a little over two times. You guys can totally help us with that."

Therefore, we'd like to know if you can do DFM or computational lithography, OPC, or even ILT and TCAD. If needed, we can invest. If investment isn't an option, in-depth cooperation and providing assistance are also possible.

Chen Yiran quickly asked Jasmine in his mind: What the heck, what are all these things?

[Optical proximity effect correction: 40nm process uses 193nm light source for photolithography, the wavelength is more than four times that of transistor gate.]

When light passes through a photomask, diffraction distorts the pattern. OPC works by pre-distorting the pattern on the photomask so that the resulting image after diffraction is exactly what you want.

GPU acceleration could be helpful, as the core computation of OPC involves matrix operations, convolutions, and iterative optimization, similar to training a neural network.

Chen Yiran was speechless and continued to ask Jasmine: "So, you mean this can be sold to Samsung to make money? Then why didn't we do this before? He seems very eager. This seems to have more potential than our EDA, right?"

There's no such thing as a more promising path; our framework is ahead of its time, and both routes offer significant advantages. However, I genuinely didn't know about OPC before, because it's a manufacturing-related concept, and my courses didn't even cover it.

What I just mentioned is something I learned at the exhibition these past two days. KLA-Tex's booth had a demonstration of the relationship between testing accuracy and OPC correction. The PDK process documents also demonstrate the role of OPC in the certification process.

Park Jae-sun continued, "Several other lines in Samsung's semiconductor business—DRAM memory, NAND flash memory, and Exynos processors—can all use OPC. Our technical team believes your framework is very suitable, but there are patent issues with commercialization. So, could you consider licensing it to us? After all, you might not be very familiar with this."

Jasmine mentally tutored Chen Yiran.

Traditionally, these fall into two separate markets. The design side is called EDA, and the manufacturing side is called computational lithography, both falling under the category of Design for Manufacturability (DFM).

However, the underlying technologies are completely shared: matrix operations, optimization algorithms, and GPU acceleration. The incremental computation and operator fusion techniques we've accumulated on the placement engine can be directly applied to OPC simulation. Samsung's enthusiasm is well-deserved; we've helped them improve their yield rate by several percentage points.

Chen Yiran: Then we can develop it ourselves, right?

We might need to hire people with a deeper understanding of the physical aspects. A computational lithography expert is essential. We also need an OPC operator at the foundry, which would save us a lot of time. An optical simulation engineer—try Amit. None of these are easy to find. Perhaps we should partner with Samsung? That would be the simplest solution.

Chen Yiran: Which one is the hardest to recruit? Can we find one domestically? Can we lure a few over with salaries?

From my very limited understanding, optical simulation engineers are relatively easy to recruit. OPC operators in contract manufacturing are much harder to find; it's a niche field, and path dependence is likely very strong.

Of course, the most difficult position is that of computational lithography experts. I boldly guess there are only a dozen or so in China; otherwise, my country's chip industry would have taken off long ago. The three major EDA companies should have more, and universities probably have some too, but they'll likely be even harder to recruit.

Chen Yiran thought for a moment and realized that the only option was to poach talent from other major EDA companies.

However, the OPC operator at the contract manufacturer has to contact the contract manufacturer directly.

Chen Yiran did some self-reflection and then said something shameless.

"Ahem, I'm fairly familiar with optical proximity correction. The core computation of OPC is matrix operations plus convolution plus iterative optimization, which is similar to training a neural network. I've always had the idea of ​​moving into OPC later."

We don't need the investment for now. However, regarding assistance, could you find an OPC operator within Samsung who wants to switch jobs and recommend them to our company?