Niagara to burst barriers of performance

By Simon London

Financial Times

Published: November 15 2004

At the heart of Sun Microsystems' attempt to innovate its way out of trouble is a radical new family of microprocessors.

"Products like this usually come from start-ups. It is striking that a company of Sun's size should be willing to pursue something so different from the rest of the industry," says Nathan Brookwood at Insight64, a semiconductor industry analyst.

Microprocessors are the central "brains" of computers. In personal computers, Intel is the dominant supplier. But in servers, the workhorses of corporate computing, there is fierce competition between Intel, Advanced Micro Devices, International Business Machines, Sun and Hitachi.

While each company claims technical superiority, until recently their processors were all based on a basic architecture dating from the early 1970s: a single number-crunching "core", designed to work on one computing task at a time, surrounded by supporting components such as memory, input and output controls.

Sun plans not just to break this mould but to smash it. In 2006, the company aims to launch a processor code-named Niagara with eight cores per chip. Moreover, each core will handle four computing tasks simultaneously - making the chip capable, in theory, of doing the work of 32 old-style processors.

If it works as advertised, this will be more than just a triumph of applied computer science. The new processor could allow customers to replace whole racks of servers with a single box containing a single chip.

Sun is not alone in seeing the limitations of single-core designs. IBM's Power4 processor, with two cores per chip, has been around since 2001. Intel and AMD plan next year to launch dual-core processors not only for servers but also for PCs.

Sun's rivals are also aware of the potential for chips that can handle two or more tasks (known in the trade as "threads") simultaneously. Intel's latest generation of Pentium processors for PCs can work on two tasks at once, a feature the company calls "hyper-threading".

But with Niagara and its successor, codenamed Rock, due to appear in 2008, Sun is taking the multi-core, multi-threaded approach to extremes.

If the potential is so great, why are other computer companies not heading in this direction? Because of corporate culture, replies David Yen, the Taiwan-born Sun veteran in charge of the programme. He argues that IBM is too conservative, while Intel has such a strong tradition of single-core design (it invented the microprocessor) that managers have been blind to the alternatives.

Whatever the reason, other companies have concluded that the extreme approach being pursued by Sun will not deliver sufficient performance gains to justify the investment.

For example, in order to squeeze eight cores on a single chip, each core will need to be simpler, slower and smaller. When it comes to sheer number-crunching ability, each of these pint-sized processing engines will be no match for the big, powerful cores at the heart of most existing designs.

"The computing paradigm is changing," responds Mr Yen. "The focus is no longer on getting a single job done as fast as possible. It is on handling complex workloads."

He makes the distinction between tasks that need to be done in sequence - where each step is dependent on the last - and those that can be done in parallel. An increasing proportion of the work carried out by corporate servers falls into the latter category. Examples include processing online transactions, handling e-mail or serving up web pages.

If he is right, Niagara could put Sun back at the front of the pack. If he is wrong, the initiative will go down as one of the great white elephants of information technology.

With every one of Sun's 1,500 or so chip designers working on Niagara-style chips and the systems they will power, the scale of investment is huge. To get a return, Niagara will have to perform substantially better than rival processors but across a wide range of applications.

Mr Yen, for one, is talking a good fight: "At this point in the development process we are very much on track. When this processor is launched in early 2006 it will not be allowed to fail."