In Depth: Supercomputers Get A Speed Boost From Specialized Chips

Math Whizzes
Hardware acceleration for general computing remains cutting-edge research rather than useful IT. "It's still an emerging technology," says Scott Studham, CIO at Oak Ridge. But the approach could benefit commercial markets in the not-too-distant future, as high-performance computing and business data processing increasingly converge. Much of today's business computing involves massive amounts of data, as marketers employ computer-taxing data mining algorithms, banks try to track global financial markets, and industrial and retail companies track streams of data pouring in from radio-frequency identification devices. "It's a red-hot topic," says Joshua Harr, CTO at Linux Networx, which sells supercomputing clusters.

In many ways, the interest in accelerators is a response to chip design trends. Specialized chips devote about 10% of their circuitry to performing mathematical floating point operations, compared with just 1% or 2% of the die area on general-purpose chips like AMD's Opteron or Intel's Xeon. Instead, those chips devote large portions of their circuits to controlling elements of the system, predicting which branch of a calculation is most likely to occur, and executing instructions based on that speculation. That leads to a balanced chip. "No matter what you throw at it, it does a reasonable job," Tokyo Tech's Matsuoka says.

Accelerator chips also run at much lower clock speeds than general-purpose chips, saving power and heat. ClearSpeed's chips, for example, run at 250 MHz. Intel's fastest Pentium chip, by comparison, runs at 3.73 GHz--nearly 15 times the frequency.

With the specialized approach, apps can be sped up by as much as multiple times their original speed, providing performance increases with less power and without the need for more networking between nodes. They excel at operations such as vector math, in which a single instruction can operate on multiple data points simultaneously, and at "Fast Fourier Transforms," which decompose signals into component frequencies and are used in many areas of science and engineering.

The big drawback so far is how difficult the accelerators are to program. "Their power is incredible, but it's hard to get at," says Alan Edelman, chief science officer at Interactive Supercomputing, which makes software accelerators for scientific apps, and a professor of applied math at MIT.

The cost of programming accelerators is "prohibitively high" for IBM customers, Turek says, not just in terms of development time, but because of a shortage of in-house skills and the need to create a new code base for specialized chips.

Easier programming of supercomputers is at the center of the Darpa program scheduled to enter its final phase this month. Under the program, called High Productivity Computing Systems, Darpa is expected to award grants of as much as $250 million to two of three competitors--IBM, Sun, and Cray--to build supercomputer prototypes that can be commercialized.

The problem in a nutshell: High-level, procedural programming languages like C, C++, Fortran, and Java are meant for programming general-purpose processors that execute instructions in serial steps, while the accelerators perform hundreds or thousands of operations at once. A few startups are trying to adapt C for FPGAs, but it's difficult to extract their performance unless you're an expert.

ClearSpeed provides an extended version of C for vector math for its ASICs. But even that approach lets programmers access maybe half the board's performance, CTO John Gustafson says. The company is working with software vendors Wolfram Research and The MathWorks, which make mathematical modeling packages used by millions of scientists, engineers, and economists, to tune those apps for ClearSpeed's chips.

Adapting accelerator technology for broad business applications could be the biggest barrier to acceptance. "You don't just worry about airflow over an aircraft," Linux Networx's Harr says. "You also need to worry about Microsoft Office and iTunes and Quake."

Illustration by Viktor Koen

Continue to the sidebar:
Four Paths To Acceleration