Total Pageviews

Monday, June 17, 2013

Supercomputing for Everyone

With the aid of the Chinese military, Intel has won itself big bragging rights: the world’s fastest supercomputer runs entirely on Intel semiconductors. It is the first time in 15 years, Intel says, that an all-Intel machine has held top honors.

More important is what this news says about computing: the kind of work done by supercomputers is increasingly applicable to the kind of work done by business. Intel doesn’t want to sell its biggest computers to researchers and the world’s armies; it wants to sell them to companies like Amazon for its Amazon Web Services.

The new computer, called a Tianhe-2, or Milkyway-2, was built at the National University of Defense Technology in Changsha, China. At its peak, it can perform at a speed of nearly 55 petaflops - with a petaflop akin to one thousand trillion instructions per second. The previous record-holder, announced last November, had a peak performance of 27.1 petaflops. Five years ago, a single petaflop machine was record-breaking.

In addition to the usual supercomputing tasks like weather analysis or geophysical research, makers of the new supercomputer also listed its capabilities for Big Data analysis. The Tianhe-2 can process 600 terabytes of data on just 1,024 of its 16,000 computing nodes.

Details on the machine were first revealed at a meeting of supercomputing specialists in late May. Its formal ranking as the world’s fastest was announced Monday at a meeting in Germany, where Intel also laid out its business vision.

Intel believes that many ordinary businesses, possibly even consumers, will soon be accessing what were once the most expensive and rarefied computers.

“The insatiable need for computing is driving this” rapid development, said Raj Hazra, the head of Intel’s high performance computing business.

Even traditional uses appear to be affected by both the power of the machines, and a new sensibility of the way problems should be addressed. Mr. Hazra noted that this much power enables atmospheric climate models to also take into account the effect of ocean behavior as well. Geologists can examine not just rock formations, but the behavior of gasses and liquids within different rocks.

This ecosystem view of data analysis has its own parallels in the commercial world. Increasingly, companies are deploying sensors across many environments to see how their products perform in the real world, or looking at mixtures of human and machine behavior to analyze things like traffic flow.

Google and NASA recently obtained a kind of superfast quantum computer to look at things like facial recognition. Lockheed Martin has also purchased one to examine complex systems.

Intel is hoping that its familiarity has an edge in this arcane world, however. Many computer engineers are already trained in Intel’s basic architecture, x86, and so could arguably move into supercomputing without having to learn the particulars of other machines. “If you force people to learn new things, your total addressable market is slowed,” Mr. Hazra said. “A world that has invested 30 years of software and knowledge shouldn’t have to learn something new.”

“We have a close relationship with all the cloud computing providers,” he said, “you could see high performance computing as a service.”

Mr. Hazra said the computer was being used for “open science,” or access by different researchers over a network. He could not say, however, whether China was also using the computer for defense purposes such as nuclear weapons research. “We are a worldwide business,” he said. “This was no different than any other business deal.”

The x86 supercomputing world is, of course, a convenient argument for Intel. It is likely that Nvidia, which contributed to last November’s record-holding supercomputer, will make its own run at the next title, however. Even more likely, according to the author of the initial report on the Tianhe-2, is the prospect that China will develop its own chips and designs.

“The Tianhe-2 is using processors from Intel, but think of the processor as a motor and they are building a racecar. They can easily swap out the motor for one of their own,” Jack Dongarra, a professor at the University of Tennessee and Oak Ridge National Lab, another supercomputing center, said in an e-mail. “Most features of the system were developed in China.”

“Intel would think the world is x86,” he said, but “the parts that go around the motor are more important.” In addition, he noted, engineers elsewhere are working on supercomputers using designs from low-power cellphone chips.