The data centers that bring us the likes of Google and Amazon are huge facilities crammed with computers that draw lots of energy. “Big Web server companies are opening up their data centers close to dams because they consume so much power,” says Anant Agarwal, an MIT electrical engineering and computer science professor who aims to change all that by putting the equivalent of a data center on a chip.
Agarwal is designing chips that contain thousands of processors, or cores. These many-core chips will save space. They’ll also use 5 to 10 times less energy than today’s chips because they’ll run at a lower clock speed, which allows operating at a lower voltage, Agarwal says.
He and his MIT colleagues have already worked out how to put dozens of cores on a chip in an energy-efficient way using a concept called tiling. That technology is being commercialized by Tilera Corporation, where Agarwal is a founder and CTO. The company produces chips with as many as 100 cores.
He and his team at MIT are pushing the technology further by looking to biology. They’re developing computers that, like living organisms, monitor and adjust themselves in real time. “We are developing architectures and software systems that are constantly observing the computation as it progresses, and constantly taking actions to achieve the user’s goals,” Agarwal says.
This new computational model, Self-Aware Computing, is the key to building thousand-core chips. These chips would be more powerful and use less energy than today’s, but would have many parts. “When you have thousands of cores on a single chip you can no longer make the assumption that everything is working and reliable at all times,” Agarwal says. “Things are going to be failing.”
But such self-monitoring, self-adjusting chips will be able to recover from problems. They’ll also be able to change the number of cores assigned to a particular computation to, for example, use more cores to get the job done faster. And they’ll be able to move computations to different groups of cores in order to manage power use.
With today’s chips, the longer you run a computation, the hotter the computer gets. Keeping computers from overheating takes energy. “Imagine if I have a computer where the longer it runs a computation the cooler it gets because it is learning about the computation and taking steps to make itself cooler,” Agarwal says, adding that that’s part of his vision for computers of the future.
Thousand-core chips are a promising route to exascale computers, which can process a mind-boggling quintillion instructions per second. A quintillion is a one with 18 zeros after it. An exascale computer would be more than 1,000 times faster than today’s most powerful supercomputer, and could simulate complex systems like the global climate. MIT is one of four research centers the Defense Advanced Research Projects Agency (DARPA) is funding to develop exascale computers.
Computers like these are probably 10 years or more down the road, but that’s right where Agarwal likes to be. “MIT has been the perfect place where thinking far out, out 10 to 20 years, is not viewed as crazy.”