DARPA JUMP Center Aims to Define the Future of Microelectronics
6/12/2018 11:14:29 AM
For such a tiny part, the transistor plays a huge role in our lives. Transistors—invented in 1947 by former ECE ILLINOIS Professor John Bardeen and two other physicists—have helped usher in the information revolution. They are ubiquitous in technology. Their low cost, flexibility, and reliability have allowed for amazing advancements in computers, machines, equipment, products—anything that involves microelectronics.
And those advancements have been incremental throughout the years. Consider this: Intel’s 22nm 3D transistors, introduced in 2011, run over 4,000 times as fast as Intel’s first microprocessor, introduced in 1971. They use about 5,000 times less energy, and their price per transistor dropped by a factor of about 50,000. The company manufactures more than 5 billion transistors every second.
That adds up to incredible speed at very affordable costs, which translates to ever-improving technology. There is only one catch, but it is a big one. It is embodied in Moore’s law, which projects that the number of transistors in a dense integrated circuit doubles every two years.
THE PROBLEM“The overarching problem is the semiconducting industry has been on a scaling path for almost 50 years,” says ECE ILLINOIS Professor Wen-Mei W Hwu, who is an affiliate of Illinois Computer Science and the Coordinated Science Lab. But he believes the pace of advancement based on this scaling process is coming to an end soon, “because as the transistors get smaller and smaller, the process has become way too expensive,” Hwu adds. In other words, you can shrink transistors only so far. Then you need to look for advancements in other ways.
This is the focus of a $3.75 million grant awarded to the University of Illinois to be part of a team conducting research to increase the performance, efficiency, and capabilities of electronics systems for both military and commercial applications. John Bardeen started a revolution in the microelectronics industry. And Hwu, Illinois Computer Science Professor Sarita V. Adve, and ECE ILLINOIS Associate Professor Nam Sung Kim, are tasked with helping to take that revolution to a higher level.
First announced earlier this year, the team's work is funded by the Defense Advanced Research Projects Agency (DARPA) and U.S. industry participants. Overall funding for the program, called JUMP (the Joint University Microelectronics Program), exceeds $150 million and has resulted in a collaborative network of six research centers spread across the country. The Illinois contingency—headed by Adve, Hwu, and Kim—comprises part of one of two “horizontal” centers for the project. Those two centers will intersect with four “vertical” centers.“We are an Applications-Driving Architectures Center,” says Kim, who is an affiliate of Illinois Computer Science and the Coordinated Science Lab. “A key role of our center is to synergistically put all the activities of other vertical centers together so we can cost-effectively produce efficient integrated systems for future computing applications. We are bringing all these pieces together to get the maximum benefit out of the individual centers’ efforts.”
The Applications Driving Architectures (ADA) Center is developing circuits and architectures to implement computation, communication, and storage applications and support the needs of the four vertical centers, which are studying RF to terahertz sensors and communication systems, distributed computing and networking, cognitive computing, and intelligent memory and storage.
The ADA Center formally began in January and is on a three-year renewable contract. Other institutions making up the ADA Center include Stanford, MIT, UC Berkeley, Michigan, Washington, Princeton, and Harvard. The horizontal centers function to:
- Drive foundational developments in a specific discipline, or a set of like-minded disciplines
- Build expertise in and around key disciplinary building blocks
- Create disruptive breakthroughs in areas of interest to JUMP sponsors
- Define a set of key metrics that they will use to benchmark and drive efforts in the defined research space
“I am a strong believer of long-term research,” Adve says. “All of my projects last many years. I tell my students it’s a story that you’re unfolding. When I think about my work over the last 20 years, I think of it as a sequence of chapters in a story where each chapter makes an impact that builds on the previous one.” To that end, JUMP is eyeing an 8 to 12-year research time frame that will lead to defense and commercial opportunities from 2025 to 2030.
THE CHALLENGESThe Illinois team—and the program in general—faces many challenges. Among them are:
- Balancing the need to build customized systems with the need to generalize those systems for greatest possible use among applications. “There are so many applications out there; we cannot design something for every application,” Adve says. Kim agrees. “Customizing hardware for different applications becomes very expensive,” he says. “So one of the objectives of our center is to provide building blocks to make a specialized process that operates in a more efficient manner.”
- Specifying software at a higher level. “The software should be synthesized so instead of people writing every line of code for each level of software, the software should be specified at the higher level—more at the algorithm level,” Hwu says.
- Increasing efficiency. “We need to think about how we can compress things into fewer operations, do less work to get higher efficiency,” Hwu says. “One of the things Nam Sung and I have been working on is lowering the system overhead by making the entire data always be in the main memory.” Data movement, Adve agrees, is highly inefficient. “Let’s say I have a self-driving car,” she says. “It has many different components. I might want to specialize the computation of each of these components, but they have to communicate data with each other. How do I connect these specialized components together so that they form a whole system with minimal overheads in communication? You need efficient interfaces.”
- Improving the interface between hardware and software. “The interface for communicating data between different components of a system is becoming a real source of inefficiency,” Adve says. “In the past, when we had general-purpose systems, everything ran on pretty much the same interface. But when you design something to be so general purpose, to be everything for everybody, it cannot be efficient for everything. We are designing the data communication interfaces and protocols for the next generation of specialized systems.”
- Working without prototypes. “The applications are changing, so system design follows a moving target,” Adve says. “It’s very hard to do this kind of research without building prototypes, but building prototypes is hard, especially in hardware..”
- Redeveloping libraries. “Software rely on libraries,” Hwu says. “These libraries need to be redeveloped and retuned for each type of hardware. Right now there’s a vacuum in the industry in terms of how library functions in the future will be able to keep up with the hardware.”
Hwu, whose career (as an undergraduate student, graduate student, and professor) spans 40 years, has seen gradual changes in computer architecture over that time.
“It has been an extremely slow evolution,” he says, “but now we’re seeing a very dramatic change. When the industry is going through slow evolution, you can take the current generation of hardware and change some of the software and do some experiments and then predict some of the benefits for making some small changes in the next generation. But when things start to be so different, you start to lose that kind of trajectory.”
In previous generations, improvements were easier, Hwu says. “You had a lot more transistors to play with and their power efficiency was getting better, so, from an architect’s point of view, we didn’t have to make dramatic changes,” he says. “Now, we need to think about how we organize the devices that we have to get better transistor energy efficiency.”
Adve agrees. “A huge part of computer performance improvement is going to come from how computer architects can organize these devices and expose them to the program,” she says. “In the past, computer architects were wildly successful in that they defined a general interface, the basic instructions that computers execute.”
That, she says, “enabled a lot of innovation in the hardware, because we knew these were the instructions we had to design to, and the hardware designers were free to innovate as long as the hardware would execute these instructions.”
TAKING THE WORK TO THE NEXT LEVEL
One of the key target applications are artificial intelligence (AI) apps, which are used extensively in society today: in medicine, finance, healthcare, education, transportation, heavy industry, the military, aviation, telecommunications, and many more industries. “All the innovations in these industries rely on computer technology,” Kim says. “We are enabling these innovations for our everyday lives.”
Hwu jokingly compares their work to that of plumbers. “We are providing the foundation, the infrastructure, for people to build safer vehicles, to build better education for students, to be able to understand financial risks better,” he says. “We won’t take a lot of the top-level application glories, but we will make sure that they can still have the glories.”
Both Hwu and Adve were involved in what Hwu calls the “previous generation of the ADA Center, C-FAR” (Center for Future Architectures Research). “A good chunk of the software synthesis that we’re doing here was also supported by C-FAR,” Hwu says.
“The C-FAR work laid the foundation for this work and increases our confidence that we can meet the challenges,” adds Adve. “In C-FAR, we were working on understanding what the issues were. Now, we are taking that work to the next level.”
“I EXPECT A BETTER WORLD”
The project involves a number of graduate students as well.
“The students get great exposure,” Adve says. “They go to these [center] meetings and interact with some of the best people in the industry and work collaboratively with them. It’s a great experience for students.”
Hwu and Kim’s students are building prototype software, hardware, and libraries. “It’s hard work,” Hwu says. “The amount of time it takes them to complete their theses is even more uncertain. So how we support our students as they take on these high-risk systems construction work that can have a lot of land mines, that keeps us thinking.”
Getting to that next level offers a lot of hope for the future of the microelectronics industry—though that future is murky at the moment, Hwu says. “It’s hard to say where we are going to be in five years,” he notes. “That’s part of the research. A lot of the future is defined in the process.”
“I think this research can be a game-changer,” Adve says. “This industry is at a point right now where the path is really not clear. This is a huge opportunity for us. It might even be a once-in-a-lifetime chance to influence in a big way where the industry goes."
“Everything that we do today will get connected by this technology. So what do I expect to see at the end of this research? I expect a better world.”