 I'm supposed to have a discussion here about the computer hardware design. We'll get to that. I just have to really tell you how much I thank you for coming here. The most impressive things is all these different path of my life coalescing in this room today. It's really impressive and moving. Thank you very much for your remarks, Perilu. It wasn't great to make me cry right before the talk, but I hope. Okay. So I have a little story, and I'll try to keep it concise, given that we had a lot of remarks already. So once upon a time, computers looked all the same, thinking a keyboard, a desktop, a display, you sit there and you got to work or do homework. Today, computers look all sorts of different things. We have wearables, we have home automation, we have virtual reality, drones. This thing, shut down your gas stove, in case you forget about it. And they can even do surgery in your belly. Why do we have all this? Because the all computing industry is driven by innovation. Innovation is a key to keep selling new products. Innovation is a key to deliver better performance, more feature, people always want more feature. I can dream, I watch Star Trek, I watch Star Wars, I can dream of application I don't have today. I know I would like to have them, somebody gives them to me. Lower power, lower cost, right? Of course cost is also another critical aspect that I keep going down to democratizing and to really push how people embrace computing. It has worked great for about 40, 50 years. It's no longer working that well. This plot that many of you probably have seen a million times is the technology node at which the top tier products were delivered for a top silicon manufacturing company in the US. You will figure out soon, even if you are not a computer designer. And so in 2001 it was 800 nanometers and every couple of years they would decrease by 0.7. So the Moore's law says we double the number of transistors every 18 months ish. That's why I give it a two years. Accommodation here. If you are not in computing, this is a little comparison node. Viruses about 100 nanometer in diameter, specifically coronavirus, like that. I didn't want to make coronaviruses everywhere. A 19 nanometer transistor, so that was 2005. One silicon atom is 20 angstrom, that's two nanometers. So everything was okay, I would say until 2014. And then things start slipping, we couldn't keep up with that two-year technology cadence. Today we are here, and what is happening is the same company has put out a product roadmap where they plan to reach seven nanometers, four nanometers, three nanometers, two nanometers, 1.8. However, those are no longer the dimension of the gate or the transistor. Those are marketing names. They say those are marketing names. Why do they do that? Because the other companies in the market has switched to marketing names. We'll give you better performance, don't worry about the size of this thing. And so, for instance, the seven nanometer that is out on the market today is exactly this product. But there's been remaining seven nanometers because the competitors call it seven nanometers, so we'll do the same. The bottom line is to tell you, we are no longer keeping up with that trend. All the benefit we got from the transistor that enable all those diversification of computing devices over the past 50 years are no longer there. And that's not the only problem. System innovation has been traditionally propelled by transistor scaling. That's waning right now. The complexity of designs has been skyrocketing. So that picture in the background there is an Apple SOC, the primary chip you have in an Apple smartphone. And you can see that from 2010 to 2018, the number of complex hardware components in that chip has just been growing. So think about managing how all these components talk to each other, packaging, everything, making sure that every piece of work is growing exponentially in complexity. And then the other problem is a meta application demands. Exactly what I was saying. I can dream today of application I cannot have. And I cannot have them not because the programmers cannot give them to me, but because the hardware behind it is not going to deliver. So this is an example of machine learning algorithms that do classification and they provide better and better accuracy, but the more accurate algorithms are the one that require more giga operations. So millions of operations per second. You can see that with the hardware I have today, if I want to ask my smartphone to tell me who are all the people in my phone, it can figure out two people a second. That's not meet and demand given the pictures I have. And it goes on and so forth for many different applications, virtual reality. We know what we want from virtual reality. Everybody knows that all of that has been there for decades, right? Why don't we get virtual reality because the hardware devices don't keep up with the latency? When you move your head, it cannot compute the new image quickly enough. It's just a hardware problem. So what has happened? Over the past five to seven years, computer architects have been coming to the rescue. They said, okay, we used to have the microprocessor. The microprocessor was always getting better because we were putting smaller transistors in it. Now that we cannot do that, let's specialize the processor. We designed a special processor for virtual reality, a special processor for graph computation, a special processor for images. So that specialization is gonna give me the improvement I cannot get from hardware. I'm gonna close the gap. A few ideas came up in the 1617. In more recent times, companies have actually produced specialized processors. So here, Google has a deep learning accelerators embedded in the Pixel 6 phone. Tesla has a self-driving chip that has two neural processor accelerators. This is a startup. It's called Samba Nova. It's actually my advisor, it's one of the co-founders, and they designed a reconfigurable neural processor. The company, to date, has received $1.1 billion in investments, and probably has given zero in sales. But the momentum is high, as you can tell. So before I go too deep in the story, I need to tell you a little bit about the ADA Center because it's the context of some of the ideas I'm gonna present today. The application-driving architecture research center was started in 2018. ADA, some people call it ADA, is now the American Disability Act. It's actually inspired by Ada Lovelace, who was the very first computer programmer in the 1800, early 1800. I want no computer, she was programming. If you ask me how she was programming, I don't know the name of that device now, on the moment, but you know, like two, the device that you can make cloth with, thank you, alum. Thanks a lot. So the center started in 2018. There are 21 faculty member, 130 PhD students from 10 different academic institutions, and there you see a whole DPI's of the center, including Tom Benish, and a number of my colleagues here. And it's funded by 12 industry sponsors, and DARPA, and NIST, and NSF, through a consortium called SRC. Now, the goals of the center, we were aware of the problem I discussed back then, was to reignite computing system innovation in the face of transistor no scaling anymore, or not at the pace we need, through sustained scalability, so new specialized architecture, embrace new silicon devices that are not the smaller CMOS transistor, but other ideas that can come up in the silicon world and non-silicon world, and then also sustained value creation. One of the key problems in the field is a thing became more and more complex. The specialization of the engineers needed to design those things got higher and higher, and because of the higher demand of specialization, the number of people that could contribute to that space start going down. I'll show you that in a second. So what we wanted to do is design flow, processes to design hardware that don't require to know everything about a clock tree, that don't require to understand every aspects of the voltage propagation through the system. Oh, sorry, these are the three themes the center were organized on. New architecture, embrace new technology, and develop new design flows. Today, we are five years in, and there is a legacy that is shaping up. First of all, we demonstrated that you can develop specialized processors in many different domains. If you look at the industry out there, it's pretty much neural network and machine learning. We demonstrated many, many different domains. We developed complete flows that can go from high level languages to a piece of hardware that run my application. And we also developed a lot of benchmark suites because we figured out that researchers, especially computer engineers, love benchmarks. When you have a benchmark, you compete to beat your colleagues and provide better solution than your colleagues. So benchmarks are lighthouses that drive the research. So this is a sample, almost complete sample of all the specialized accelerators that were developed in the Ada Center. Graph processing, privacy, genomics, 3D image processing, speech recognition, also next generation machine learning, so natural language processing. We're also developing a specialized accelerator to support the explainable AI algorithms, which is becoming, the explainability aspect is becoming pretty required as AI gets deployed in very, in domains that are affecting human very closely. So the air out design specialization has become. The challenge is now that we don't just have one processor or maybe two designed by two companies around the world and leveraging the small and transistors. Who's gonna design all this stuff? So I went to the Bureau of Labor Statistics. These are the number of software developers reported by the Bureau. It was about a million people in the US in 2012. We're now over a million and a half. And the projection is that by 2031, we're gonna have over two million people in this country who are software developers. These are system software programmers. Those individuals that can program the lower level of the operating system, the drivers, the connection between hardware and software. And these are my computer hardware engineers, the people in charge of designing all those specialized processors. I know you cannot see the number because the pile is low. So let me bring it up for you. It's 81,000 in 2031, 77,000 today. That's one to 20. So we have one hardware designer for 20 software developers. And the pressing thing is this went up significantly. This went down. And I think part of the reason it went down is the level of specialization required is increasing, continuously increasing. So we have a problem. I'm sure you got that by now. I'm gonna tell you a little bit how design works today and how we propose to solve the problem. So that's a wall, software on one side, hardware on the other side. How do I design hardware? I come up with a schematic. I develop an accelerator, a chip that is very good at doing one thing. And then from the architecture design, I have to do the view the side design, place and route, develop a prototype, embrace a new technology if I can. Great. On the other side, I have the software programmers. They pick up their favorite programming language. Usually today, people don't necessarily have to program in C and C++. There are these domain-specific languages that are super powerful for the domain you're programming in. Like Heli, it's a graphic programming language. It has primitives that like you do image transformation with one line of code. So the productivity of the software developers much higher, many fewer line of code to describe what they wanna do. And then that gets related to C and C++ because that's what we have out there. All these domain-specific languages that just compile into C and C++. Lots of people doing that job. Very few people doing this job. And how do we get this C and C++, which is very good at running on your genetic processor, how do we get them to run on this much more higher performance specialized processor? Well, we need a uniquely talented engineer that can understand the software side and the hardware side and change the compiler so we embrace the specialized hardware that we have on the other side. That's today. This is what I'm working on and I hope we can do. One, create design flaws that are easy to adopt. Design flaws that go from the high level domain-specific language to a piece of hardware without having to understand everything and everything else. And then increase the participation in the hardware design field so we're more blue people to design more components. Let's talk about ones. First of all, this thing about these design flaws, we did it once before. When I say we, I mean she. The she is Lynn Conway, who also is the namesake of our department chair. So let me tell you this story very briefly. In 1971, Intel released their first microprocessor. It was a 2,300 transistor drawing where every transistor was 10 micrometers. How did they do it? They sat at the table with a big piece of paper and they start drawing where every transistor goes. And if they do a mistake, they erase and they redraw. It was all done by hand, pencil and paper. And they did that and then they figured out a better way to do a processor and a better way to put more transistor. So by 1978, they were doing this. 29,000 transistors, very difficult to track on a piece of paper. So they were at a point where it was very, very difficult to continue to do better chips not because people didn't know how to scale a transistor but because it was so hard to track 29 pieces of thinking all on a piece of paper. So Lynn Conway came around. She's our colleague by the way. She's our emerita professor. And she came up with algorithm and rules that could automate how I lay out the transistor on a chip. And she came up with software program that help you not violate any of the rules so that people could write something that look almost like C. It's not a C, but it's much more manageable than a drawing and it's much more editable than a drawing. And not only she did that, but then she started teaching classes. And she went to different university teaching a class on how to do VLSI design. So there were kids in their 20s who could design microchips comparable to the one that the engineers at Intel were swiping and spending all their time trying to keep the story straight. That simple, my simple wasn't, but the innovative change radically changed how people were doing silicon design and open up a whole new scale or complexity that became available. In much smaller scale, the ETA Center has also been trying to work on creating infrastructure that connects these high level languages with this specialized microprocessor there. And how did they do that? We have a group of 21 different faculty, everybody has idea in a different direction, we get together, we discuss idea, we compare them, but as faculty member we love our ideas the best. So we have a few ideas here. This is approached by three colleagues from the center, one from Princeton, one from University of Washington and one from Harvard. And what they figure out is how to compile several programming languages for machine learning into accelerators, but these are general approaches, but these are four different accelerators for machine learning. And the way they do it is they create the control flow, the program in the high level language, we create small programs of everything that the accelerator can do. And then they use this ILA language to prove the equivalence between the primitives and the small program. And they match, then we have a match and we can compile the program into the accelerator. This is another project from one of my students who is here today. And basically the idea is I have my high level language, it has all these powerful primitives, I have a hardware accelerator or many of them who provide some function at very high performance. And what I do is I have an infrastructure that does trial and error and figure out what's the best match. So I know I have a match, but some matches don't give me a performance benefit so I don't need them. And you can identify the best match among a list of multiple processors. So here we show that we took the graphic programming language, it's used to express graph algorithms, it's very efficient for that. Two different graph accelerators. And if I only use one, I get a 45% performance improvement. If I use only the other, 17% is by hand, I pick the best for every call, I get 58% and with Nikola's automated way, I get 57%. So I can get very close to optimal without necessarily, without at all having any manual managing of that. These are other two infrastructure developed by PIs in the center to help computer architects develop silicon prototypes, silicon in hand prototype very quickly. They give you the motherboard, they tell you how to do the pinning, they have libraries for everything so that as long as you start with the description of your hardware in RTL, and you need $50,000 for the chip manufacturing, you can very quickly have your prototype. One, my colleague Mike Taylor told me that one time some of his colleagues at the University of Washington said, I have this special application, can you see if you can put it in one of your chips? And they went to look and the application was in Python. And he said, well, start by converting to C++, you may get what you want right there. I can't use increasing participation in the hardware design field. So we need more people as you saw that. And all of you already, well, all of you in the front line have already spoke about some of these initiatives. One is Africa Undergraduate Research Adventure, which is an undergraduate summer research experience for undergraduate students from Africa. I launched that work with Todd Austin in 2019. And so think about 2019. So we have two years of pandemic and two years of in-person. But as of today, we already see that the students who have graduated and have taken this program, they are very likely to pursue graduate studies. So over 60% of the participants are in graduate school, the vast majority in the US and the vast majority of those at the University of Michigan. So it's between 50 and 75% at U of M. But we're not selecting anyone to pursue graduate studies, whatever they want. This program has been generally supported by Ale Gallimore for many years. But we also have some industry sponsor who are very interested in getting more hardware designer and the Buffett Foundation. The other program is computing chaos. As Mary Lou mentioned, I started with her and Amir Kamil in 2015. Amir Kamil came a little bit later, maybe 2017. I don't know if you were our colleagues in 2015. Okay, good. And the idea is to have intervention in the classroom, in the lower division programming classroom to foster a more inclusive climate. So we teach them about stereotype threat. We teach them about implicit bias and we give them assignments that help them reflect about these aspects. And then the other thing we do is, we do monthly workshops that try to open their mind to opportunity they have by studying computer science that are beyond sitting in a cubicle and program the rest of your life. And also we have specific trainings that we do for the GSI and IAS of all those classes. So today we have trained 799 GSI and IAS and we have impacted over 70,000 students because computer science classes are very large. So each semester we have several thousand. This program has been supported at the third center initiative and by SIGWIT. So looking ahead, we are at an inflection point right now. As you probably have heard from the news, there have been several challenges in the semiconductor supply chain for many reasons, but the cost of fabricating chips, the ability to fabricate enough chips, the trust of chips produced by other companies and the leadership of the United States in this space. This plot shows you how many companies were delivering top node silicon in the years in the column. So in 2002, there were 25 companies in the world that would give you the smallest technology node and they would produce chips at that node. As time has gone by and the transistor got smaller and smaller, the number of players in the industry have shrunk. Also, this number in orange is how much does it cost to set up a fab, to set up a plant that can produce those silicon? So it used to be about $2 billion and today we're talking about $20 billion. So pretty much we have left with three companies, only one is in the US. So now you can solve the puzzle on the first slide. And these are all competing to all three of them stay in the market, which I actually think it's a great thing. They all switch to marketing names for that technology nodes and the capital investment to even start something like this is unapproachable by anybody who's not already in the market today. So it's only the survival of the fittest. The United States government in response to realizing the situation, the pandemic really created a exacerbation of the supply chain challenges as long as CHIPAC last August with the goal of bringing back the leadership in semiconductor into the US, growing the population of Americans and engineers that work in the field and they also put $52 billion on the table. As a result of this, a number of companies have already announced that they're gonna build fabs in the US. This started for the past couple of years. As you can see, I give you here the name of the company where they're planning to launch their fab, how much is estimated or how much they committed to invest. We're all talking at the $20 billion a piece here. And so as you can see Intel, TSMC, Samsung, all the free major player, Micron is a memory company and Texas Instruments and Goblet Fund, they are not necessarily doing the frontier node but they're also investing in growing the production of microchips. Here at the University of Michigan, I started to work with Eric Mikkelson to gather our strengths and our direction and propose the opportunity that we can provide, the problems we can solve as Michigan Engineering. So we launched this very recently. We launched this Michigan Advanced Vision for Education and Research in Integrated Circuit, MAVERIC. This is a retreat we did with 50 colleagues about a month ago. We identified several aspects in which we can provide innovation. For instance, what if we can manufacture transistor without rare earth material? So we wouldn't be dependent on the few country who have that material. What if we can provide trust in silicon even if we didn't manufacture it in a trusted fab? This is our Lurinanum Fabrication Facility and I really hope in the next few months we will be able to launch something with government funding in this space. Yeah, I'm not looking at you Alec. Don't worry. So I just want to thank a few people. The AB Research Lab, that's my students. It's called AB because Todd Austin is an A, I'm a B and he's Austin Bartacco, we meet together very often. They're all looking at me here. I just got myself out of this picture. And then this is my VPAL team, my Vice Provost for Engage Learning team, who also has been very supportive and I'm so happy to work with them. They're the best part of my job. And my family, my kids for putting up with all the work chat that happens at dinner and Todd for driving, cooking and everything they need for them. So thank you very much. Oh, wait, wait. And Mary Lou for keeping me alive all these years.