Chip redesign may herald a new dawn

Marc Tremblay wants to serve up a grid computing network on a single computer chip. If he and his team at Sun Microsystems are successful computing will arrive in a new future where individual computers will be processing titans, even outpacing many grid systems.

Grids are networks that link together a number of computers into a powerful team , forming an ad hoc supercomputer - a versatile, muscular system that can efficiently power through the most complex and demanding computing tasks.

Grids are hot in the world of computing right now.But imagine a grid on a chip - in essence, placing a number of tiny computers on each individual computer chip, giving a single computer the low cost, high octane computing capability of a grid.

That vision - one which fundamentally changes the way in which chips are structured and tackles head on some of microprocessor design's biggest challenges - entrances Dr Tremblay, vice-president and chief architect, processor and network products, and Sun Fellow.

The initiative, called Throughput Computing, is so central at Sun that it now occupies 1,300 of Sun's 1,600 chip engineers. "Companies have been building microprocessors for the desktop," he says."But we're seeing a fundamental difference between desktop computing and server-side computing."

That difference - which centres on the kinds of tasks a computer is asked to do, and how it can do it most successfully - emphasises the key drawback to today's chips. Most home computer users want to do one process efficiently - say, create a document or make a spreadsheet.

Chips are designed to focus on getting through that one task quickly and smoothly.

But researchers, computer animators, scientists, designers and others who need the power of computer servers to do several computing tasks at once find the capability of their computing systems are held in check by chips focused on doing one thing well, rather than a number of things quickly at the same time, he says.

This issue is further complicated by another chip design shortcoming. Today's chips are like Porsche engines placed in a Lada - the engine can do far more than the system that it runs.

The bottleneck comes in the memory department.The microprocessor is fully capable of initiating a number of tasks, but then, it must waitfor memory chips to go get the data needed to complete the task. The gap is dead time for the microprocessor, on the order of the infinitesimally small - only as long as 100 nanoseconds. But such delays, multiplied by the millions of tasks a computer carries out, turn into significant slowdowns for processing-heavy jobs.

For a scientist, those delays turn into hours of extra time needed to analyse data or to run a simulation. For an internet user, those tiny time outs mean waiting seconds longer for a page to download, a Google search to end, or a purchase to complete on a share trading site.

"What would be nicer is if you could process 10 times more transactions," says Dr Tremblay. You might think the solution is to create a chip with even more transistors (the tiny processing engines of a chip) on it, but you'd be wrong,he says.

In the past, adding transistors to chips has guaranteed the durability of Moore's famous law, that chip power will double every 18 months. Intel's latest Itanium chip has an extraordinary 410 million transistors on a surface the size of a stamp.But chip designers have recognised for several years that Moore's Law will expire if new ways aren't found of structuring and building chips.

The Lada system has to be wholly redesigned, rather than putting ever more high performance engines into the car.

"The biggest issue with Moore's Law today is not really putting more transistors on a chip," Dr Tremblay says. Even with ever more transistors, "we're seeing a slowing downon microchips. We've mostly explored the real estate made available to us through Moore's Law."

Sun's answer is to look to a new chip design to tackle the memory issue and the problem with existing resources being underutilised. He acknowledges that Sun is "taking a pretty drastic approach to it" - totally rethinking how a chip operates.

Sun is dividing the chip into separate processing regions - a "core", or tiny microprocessor within a microprocessor - each with its own memory chips. Each core can handle four "threads": "a thread could be an application, or one Google request," he says.

With each core focused on its own four threads, it can operate more efficiently - data from memory needed for the first thread will be coming through as the third or fourth thread puts out their own requests to memory, explains Dr Tremblay.

Sun says that using only 10 to 15 per cent of the available transistors, a single core would achieve40 to 50 per cent of the performance of one of today's chips. Sun plans to offer two, four and eight core versions of the chip, with a preliminarytwo-core chip appearing later this year.

But managing memory and cores in this way creates a number of problems. A key issue is the amount of power needed to run the memory withoutcausing overheating. Memory also takes up valuable space on the chip.

"This is the art of creating a chip. There are trade-offs," says Dr Tremblay.In addition, the operating system itself will need to be designed to work in an elegant ballet with the special chips.

Not all chip manufacturers believe such a total rethink is necessary. Intel is approaching the problem differently - keeping as many processing tasks within the microprocessor as possible to curtail the use of chip memory.

But Dr Tremblay is convinced that core-based chips are computing's - and Sun's - future. Sun predicts that chips in 2006 will offer 15 times the performance of today's processors. "We're going to get there first and fastest," he predicts, "but we expect everyone to go there eventually."