Arijit Raychowdhury, an expert in digital and mixed-signal design told that packaging is an area that process engineers must understand...
The new frontier of leading-edge IC design is packaging, according to Arijit Raychowdhury, an expert in digital and mixed-signal design who teaches VLSI courses at the Georgia Institute of Technology and who received the 2018 IEEE/ACM “Innovator Under 40 Award” at that year’s Design Automation Conference.
Integration is the holy grail of VLSI design. Next-generation IC design lives or dies on the next feasible advance in integration. Historically, the VLSI community has depended on progress in process node technology to overcome the next fast-approaching bump on the road toward higher integration. But times are changing.
The chip industry understands that scaling in accordance with Moore’s Law has been slowing. The industry appears reluctant to confront an imminent upheaval in chip integration — an unmistakable transition from process to packaging technologies.
Raychowdhury is one of a rarified few who have zeroed in on this trend. Citing Advanced Micro Devices’ Zen and Intel’s Lakefield processors, he told EE Times Europe that packaging is an area that process engineers must understand. Intel is using an advanced package integration technology called Foveros for its Lakefield chips, Raychowdhury noted. “AMD is using similar integration technology to combine 7-nm CPU with 10-nm I/O using package-level integration. It helps improve system yield for them. Their flagship product for this is Zen 2.”
Others are notably taking similar approaches for package-level integration. “Apple, Qualcomm are all pursuing this, and they are at different levels of maturity,” said Raychowdhury. Put simply, leading CPU vendors looking to the future are “all about heterogeneous integration of the package.”
The emphasis on leading-edge IC design is shifting from the process to the packaging. The problem is that “there is very little understanding, at least in the U.S., about how this transition is going to happen,” according to Raychowdhury. He singled out TSMC as a company “doing a better job” of letting the industry know “what is on-die versus what is on-package.”
During our interview, Raychowdhury stressed the importance of “connecting the dots” in the engineering world. Linking theories learned at school with designs in the real world is one obvious example. Knowing how the business of technologies connects with Wall Street valuation is another. While technical papers submitted at ISSCC or VLSI Symposia inform engineers about cutting-edge technology, there is a huge middle ground that must be explained between advanced tech papers and books that teach engineering basics.
Our conversation covered challenges in teaching EE courses in the pandemic era, transistor scaling, new fields of expertise for next-generation VLSI designers to master, and what sort of engineering students are most likely to succeed in the real world.
EE Times: I see you’re at home. Is school still in session?
Arijit Raychowdhury: We moved to online classes in the middle of March. We have students from all over the place — you know, from China, from Vietnam — so it’s crazy to hear from a student from the West Coast saying, “Oh, it’s a lot of work.”
EETE: How do you teach online?
Raychowdhury: We video-record the lectures ahead of time and post them online. Then during normal class hours, we just log into Blue Jeans, Microsoft Teams, or some other video conference system. Then we do a little more of a Q&A. I answer students’ questions and all that, as opposed to going through the entire lecture. It has been OK. I mean, I wouldn’t say it’s that bad, but it’s been OK.
But I think now you’re seeing lots of students who don’t have proper internet connections. This could be a struggle even within this country right here. There are students in rural areas, and they don’t have a stable internet. And you know, we already had tornadoes a couple of weeks back. And there were students in southern Georgia whose houses suffered some damage and power outages. It’s just a mess.
EETE: So despite all the advances in technology, we are still at the mercy of nature. Is that right?
Raychowdhury: In all sorts of ways. Yes.
EETE: Over here at AspenCore, we are on a mission to discern a gap between what our readers thought they’d already learned in school and what they have to do on the job. From your viewpoint as an instructor, are you getting feedback from students such as, “Gosh, I wish I’d learned that in school”?
Raychowdhury: That’s one thing students always mention, and this can be very specific to certain instructors as well. Students are always interested in understanding how things they’ve learned in class apply to the real world — which products use which fundamental techniques and [apply] what they’ve been taught. This is what we hear often in academia.
Some academics have moved to academia from the industry. They have a better understanding of what’s going on. I think students — undergraduates particularly — enjoy taking courses that have more implications [in terms of] on what the industry is doing and whether they can get jobs. Then there are some courses that are fundamental; we can’t stop teaching them Laplace or Fourier transform.
EETE: So these must be taught.
Raychowdhury: Right. But [students] sometimes find it hard to connect them to real-world applications. They are always looking for how the theory connects to practice.
You asked me about my connections with [EE Times Europe sister site] EDN. I’ve been reading EDN since, I don’t know, 2001. I used to work at Texas Instruments long back before I went to grad school. So, the first design that I did at TI actually got this EDN Innovation Award. It was back in 2003. It wasn’t just me; our whole team got that award. Because it was the first award I actually got in my life, that was cool.
EETE: Speaking of EDN stories, one of the most popular articles is about “setup and hold time.” Consistently, the basics of setup and hold time have been very popular. So where could we get more of that kind of stuff out there?
Raychowdhury: I think that’s a perfect question. I agree. These days, in our VLSI course, I spend about 20% of my time just talking about setup and hold time. I did not understand how complex that was before I started working in the industry. And if you really want to understand what hold time is, and if you design a flip flop and how you clock the two stages of a flip flop so that you can get the best hold time … these are the things that you only learn by building it and seeing how it works and then deploying it in millions and billions. We need to get someone to write about it. I think you need someone who, first, [knows] the theory but also understands what’s going on in practice. This is exactly the kind of thing missing in books. It’s missing in most articles as well.
EETE: In the academic world, it takes time for a book to be published. It takes time for a course to be developed to cover a technology. Are there things you see that you wish you had an opportunity to create a course for?
Raychowdhury: For example, look at technology today, and people are saying that transistor scaling probably is coming to an end, whatever it means. Technologically speaking, maybe scaling isn’t going as fast as we would want. But in terms of memory technologies, for example, or back-end-of-line transistor technologies, I think there are lots of new things happening. The industry is moving very fast in that particular domain. And there are no good books, because these are all the black magic that the industry talks about.
Another trend that I see, which is not covered in books as well as it should be, is integration — 2D, 2.5D, and 3D. We all wonder: What does it mean, then, and what are the different implications?
But [discussions about it in the industry are] optimistic. They talk about layers and layers of transistors, which is never going to happen. Even if it’s technologically feasible, it’s just economically not feasible. Even for the short-term means of getting there — like the chiplet technologies that Intel has been working on. I don’t think there are good books on [that].
EETE: Are there emerging technologies that we ought to be highlighting or we should be bringing to people’s attention or getting them prepared to be able to use? There’s a lot of excitement about things like quantum computing, but in reality, quantum computing is a decade away. [Are there] things that are a little closer so that we might say, “Hey, you know, in the next year or two, you might be called upon to work in this area, and here’s some background information to get you started”?
Raychowdhury: I think one of those areas would be that process engineers now need to understand packaging. The way I look at that trend now, if you look at what AMD is doing with the Zen processor or what Intel is doing with Lakefield, you see it’s all about heterogeneous integration of the package. A lot of back-end engineers in these companies are now going to use their skills to build dense packages. I can see that happening.
Many of our research programs [at Georgia Tech] are industry-funded. So I can see that those in the industry are asking us to look at these kinds of things — work from a design perspective as well as from a processing perspective. There is very little understanding, at least in the U.S., about how this transition is going to happen. TSMC is probably doing a better job of making sure that people can understand that this is a smooth transition between what is on-die versus what is on-package. And I think this would be one area where I would suspect people need to be ready very soon.
EETE: As a professor, when you see a student start work at TI, for example, what general knowledge do you think he or she should have? What are the basics — other than a personal specialty — that he or she should absolutely know or seek to know? Any advice?
Raychowdhury: I feel like people who do well in the industry eventually have a broad understanding. Today, I tell my students to take courses all across the board, not just courses in your area of research. A good example would be people who are doing process technologies. They need to understand physics and chemistry really well.
Device people really need to understand materials. If they don’t, it’s a problem. It’s harder for materials people to understand devices and technology. I think understanding how basic chemistry works helps. Because as we change the paradigm, we are looking at new materials.
Similarly, I feel like a lot of circuit designers, particularly analog designers, who go into the industry do not have or have forgotten basic math. So I think they have to relearn those things once they go to TI or whatever company.
Sometimes they have the skill sets, but they do not necessarily have a broader, general understanding of engineering — including math, physics, and chemistry, which are often required.
EETE: When you showed up at the doorstep at TI, you were the young guy. Did you already have a Ph.D. at that time?
EETE: So when you showed up, what things did you wish you had studied before joining the company?
Raychowdhury: I had all the background knowledge, but what I was missing was, I didn’t know how to connect the dots. How is this particular subject material connected to that subject material?
I think that’s something that you don’t learn in college, because you’re learning all these courses in separate semesters. At some point, all of these things need to click together. My first six or eight months at TI were essentially spent trying to understand how these different components are connected.
Universities now are trying to do more hands-on design-based courses in the senior year and even for graduate students, which I think will help a lot.Some people have the natural ability to hack things and write software and break things and put them back together. I think they are better engineers. They will be able to connect better when they enter the workforce. So I think more vertically integrated projects and courses are important.
EETE: As you develop new courses for the next year, what are the things that you think will be must-teach topics?
Raychowdhury: I teach circuit courses. From a circuits perspective, I think a lot of students have gone down to the deep end. They’re just interested in, you know, understanding how EDA tools work. Like, how do you know the different new options that Synopsys has come up with or new back-end tools Cadence is producing? And they want to just do projects and get the skill set.
I’m actually taking the fall semester off and trying to build a new course I can use in the spring when I go back. I want to teach the fundamentals. When grad students — and I mostly teach graduate students — graduate without having a good understanding of signal and noise, that’s a problem. And these are circuit designers.
In the VLSI space, we don’t have a lot of good fundamental courses. That’s mostly because of demand from the students. Students mostly want things that can help their skill set. They don’t really understand that it doesn’t take long to pick up those skills after you graduate. They need more of a fundamental understanding.
If you look at the VLSI curriculum — and this is common across the entire country — none of the schools are teaching fundamental VLSI circuit design.