May 10, 2011

Apple’s iPad Gets a Linpack Workout

Michael Feldman

The popular Apple iPad is certainly no supercomputer by today’s standards, but if you could transport one back in time 25 years, it would be one of the fastest machines on the planet. According to a recent article by John Markoff of the New York Times, the performance of today’s iPad 2 would rival that of a 1985-era Cray 2 machine.

University of Tennessee’s Jack Dongarra has been running Linpack on the iPad and found the little dual-core wonder tablet does quite well with the linear algebra benchmark. That’s a little bit surprising considering it uses an ARM-based chip and is certainly not optimized for floating-point performance.

Markoff writes:

To date, the researchers have run the test on only one of the iPad microprocessor’s two processing cores. When they finish their project, though, Dr. Dongarra estimates that the iPad 2 will have a Linpack benchmark of between 1.5 and 1.65 gigaflops (billions of floating-point, or mathematical, operations per second). That would have insured that the iPad 2 could have stayed on the list of the world’s fastest supercomputers through 1994.

The Cray-2 was a custom-built vector supercomputing that delivered 1.9 peak gigaflops in the 8-processor version, and unlike the iPad, certainly was built for heavy-duty number crunching. But the Cray-2 took up a small room and needed to be immersed in a special refrigerant called Flourinert to keep the machine cool.

Apparently Dongarra and his team have been tossing around the idea of building an iPad cluster using a couple of stacks of the tablets. Unfortunately, the iPad is basically a closed system, so each one would have to be hacked into in order to hook them together via the built-in wireless communication.

Plus at $400 a pop, that’s going to make for a pretty expensive cluster, price-performance wise. But Dongarra thinks the low power consumption (and the fact that it runs on batteries) is the compelling part. Of course, the whole idea of a mobile supercomputer is pretty interesting too, especially if the cluster software can take into account nodes which can come and go as they please.

Full story at The New York Times