Hide menu

LiU research behind Intel’s superchip

"Polaris" may be the first step towards the next generation of high-speed, low-power computers: a chip with 80 cores, performing 1 Teraflops while consuming no more energy than an ordinary light bulb. Behind this great progress lies research at Linköping University.

"The building blocks we designed in Linköping make up the base for the Teraflop chip," says Sriram Vangal, an Intel researcher and a doctoral fellow at LiU’s Division of Electronic Devices.

The blocks are floating point processors and compact routers that efficiently direct the traffic within the chip. Each one of the 80 cores contains two such 32-bit processors, a router and a memory unit.

The work is described in Vangal’s 60-page licentiate thesis 2006. His research results together with his supervisor professor Atila Alvandpour have also been published in several scientific papers.

"When I came back to Intel I tried our theories in practice and they turned out to work perfectly well," says Vangal.

The chip, so far aimed for further research, was built at Intel Circuit Research Lab in Hillsboro, Oregon, by a group led by Sriram Vangal.

"We were a very small team, less than a dozen researchers, starting in August 2004 and finishing in June 2006. We had two years of back-breaking work," he says.

In November 2006 the completed silicon components were delivered from the Intel plant in Ireland. Still, after about a thousand test runs not a single bug has appeared.

One Teraflops means a trillion (10 raised to 12) floating point operations per second, a task “Polaris” performs with a power consumption of 62W – a world record! All components are squeezed in on a chip the size of a fingernail. The first time the teraflop wall broke was eleven years ago – in a supercomputer with 10 000 processors, consuming 500 kW!

With the new technique, scaling of PC’s to Terabyte level will be possible, providing desktop computers for artificial intelligence, speech recognition, HD-quality video communication, photo-realistic games and multi-media data-mining.

Revolutionary, in the concept, is the architecture with 80 relatively simple, 3 sq mm communicating cores, tightly packed in a two-dimensional mesh. Sriram Vangal compares them to Lego building blocks and since all are equally large and have the same features, the construction is fast and simple.

"You create one such ‘Lego tile’ and make as many copies you like. Then you can put all the tiles together in one single die," Vangal says.

Today’s most advanced processors contain two or four cores. So why 80? One advantage is that you only use the cores needed at each occasion. Those not in use stay asleep in order to save power. The large number of similar cores also enables parallel operations and thus even more advanced applications.

Still, a lot of work has to be done. So far the Teraflop chip lacks the properties needed for booting an operative system, running applications or creating a user interface.

Åke Hjelm 2007-08-13

Page manager: therese.winder@liu.se
Last updated: 2009-06-03