Back in the late 1990s, Justin Rattner got a special sense of satisfaction every time he drove by a nondescript Intel building in Beaverton, Oregon. Inside, researchers from Intel and Sandia National Labs were assembling the ASCI Red supercomputer, the first computer capable of doing one trillion calculations per second.
“When Chuck Yeager cracked the sound barrier or Armstrong landed on the moon, I wonder if they had the same feeling,” remembered Rattner, Intel’s chief technology officer, in a 2006 news release on ASCI Red’s shuttering.
Now Intel says that it can put the processing power of ASCI Red in the palm of your hand. Literally.
Intel does this with a new chip, code-named Knights Corner. Knights Corner crams more than 50 general-purpose Pentium microprocessor cores onto a single chip. All by itself, Knights Corner can perform about 1 trillion mathematical calculations per second. In 1996, it took 72 cabinets of servers for ASCI Red to pull off the same feat.
That’s not bad for a chip that just a few years ago seemed to be a failure. Knights Corner was built from the ashes of Intel’s failed graphical processing unit (GPU), called Larrabee.
Larrabee didn’t work out, and rival Nvidia reigns supreme in the GPU market. But Intel hopes to give Nvidia a run for its money in high-performance computing, a place where Nvidia has been making inroads with its Tesla processors.
The Tesla chips can do a lot of calculations without burning up too much power, and in the past few years that’s won them some fans in the supercomputing set.
Back in June, 17 of the world’s top 500 supercomputers used these graphical processing units. Now that number has jumped to 39, according to Erich Strohmaier, a senior scientist with Lawrence Berkeley National Labs who helps compile the Top500 list. “I see more and more talks given where people figure out smart ways to use them,” he says. “They offer a way to assemble a lot of performance without blowing the power envelopes that you have.”
Today’s supercomputers use a combination of Nvidia chips and x86 processors, so the scientists who use them have to write special code that is offloaded to the Nvidia GPU and run there.
With Knights Core, they won’t necessarily have to rewrite their code, says Karl W. Schulz, associate director for application collaboration with the University of Texas at Austin’s Texas Advanced Computing Center. “With Knight’s Corner, the programming model that you’re using is the same programming model that you’re using on an AMD or a Xeon,” he says. “You get good parallelism right out of the box, which is convenient.”
Schulz should know. He is building a massive 10 petaflop (10 thousand trillion calculations per second) supercomputer, called Stampede, out of these next-generation Intel chips. To do this, his team takes server cards that have a Xeon and Knights Corner processor and slides them into specially designed 4U (7 inches tall) Dell server boxes. They expect to get 8 petaflops of performance from the Knights Corner chips and another 2 from the Xeons when Stampede goes live. The 4U servers hold just one card now, but they are designed to eventually hold two, which means at Stampede could double its power.
When it goes online in January 2013, Stampede will have 10,000 times the processing power of ASCI Red.
“It’s amazing,” Schulz says of the rapid advances in number-crunching chips. “We’ve had cases where we’ve deployed something that filled an entire supercomputing room, and then a few years later you deploy something in only one rack and it’s more powerful.”
James Reinders spent two years working on ASCI Red. A parallel programming evangelist with Intel, he says that he gets emotional when he holds a Knights Corner chip in his hand. That’s an unusual experience for an engineer.
“It’s a reminder of how fast this industry moves,” he says. “I spent a lot of my life with a lot of my co-workers designing ASCI Red … to think that I can hold that in my hand now, it’s humbling.”