Chipping away at processors

Internet appliances - lightweight, mobile, low-cost and full-featured - are poised to come on strong. According to a report earlier this month from IDC, the new consumer devices will surpass PCs in the U.S. in 2002. In that year, according to IDC, there will be 23 million desktop PCs in American households - but 25 million Internet appliances, including Internet game players, Internet TVs, hand-held computers, Web slates, e-mail terminals and IP screen phones.

That burgeoning growth will be reflected globally, the IDC report shows. In 1999, there were 11 million Internet devices worldwide; by 2004, there will be 89 million.

Prospects like that almost justify the ballyhoo surrounding the debut in January of an entirely new class of chip for mobile devices, the Crusoe line from Transmeta. With the panache of a Hollywood premiere, Transmeta President and CEO David Ditzel shared a dais with Silicon Valley luminaries such as Linus Torvalds, developer of the Linux open operating system, and Dave Taylor, inventor of the computer game Quake.

Transmeta has built a new kind of chip, transforming the way instructions are processed rather than just doing the job faster and more powerfully, Ditzel said. "In human terms, what we've done is build a machine that responds well to human interaction," he said in January. "What we're saying is that if it's mobile and it has a browser, it will use a Transmeta chip."

"That may not be hyperbole," said Donald Harriman, a technology analyst with Lavin Associates. With the TM5400 for laptop PCs and the TM3120 for Web devices, "there's at least a chance that Transmeta's chips may be the enabling step that helps mobile computing take off," he said

To understand why Crusoe rates all this attention, it's necessary to see what it does differently.

With most current chips, a programmer writes a piece of software in a high-level programming language such as C++. It then goes into a second software element called a compiler. This translates the software into a sequence of instructions for a specific microprocessor, such as one from the Intel x86 family or the clone chip from Advanced Micro Devices.

That's tough enough. But when new chips come down the line, they must run as fast as possible and be backward-compatible with software written for earlier chips. Most standard chips today solve the speed problem by breaking a sequence of instructions into smaller sub-sequences and running them in parallel or out of order.

The cost of building in this speed and compatibility is a large increase in the number of transistors on each chip, which makes them consume much larger amounts of power than their silicon forebears. That's particularly bad news for devices intended to go mobile because battery designs have plateaued.

The new Transmeta chip is specifically constructed to execute sub-sequences more efficiently. The translation of x86 instructions into sub-sequences - and all the ensuing speed-up tricks of re-ordering those instructions - takes place in a special layer of software that sits between the chip and the program. This code-morphing software dupes the program into thinking it's being run on a standard compatible processor, translating the x86 instructions into "very long instruction word" instructions up to 128 bits wide. Intel chips are limited to 32-bit wide sets.

Doing the translation in software allows more complex optimization of the system. For example, once the Crusoe chip has broken a piece of code into sub-sequences, it can store that translation in a buffer and recall it if needed. A hardware approach requires the chip to do the same translation every time the sub-sequence crops up.

And if Crusoe finds it's running parts of that code often, it will look for ways to optimize these "high repeat-rate" instructions even further. So the more often the program runs, the faster it can get.

On the compatibility front, Crusoe can imitate any microprocessor and can do several different types of processing at once. That means that Transmeta can write new code-morphing software to keep up with new hardware technologies as they come along.

Whether this re-design will interest consumers depends on their susceptibility to geek allure. But the end result should appeal to everyone who wants to access the Web from a moveable device - which, most analysts say, will soon be all of us. Laptops built with the Crusoe chip should consume less power, allowing users to go longer without recharging their batteries. For example, a Crusoe laptop should be able to double the run of the average laptop battery, which is currently about two to three hours.

"We expect that this product will go all day," Ditzel said.

And by using fewer transistors to do its translation job, Crusoe runs cooler, eliminating the need for a fan to cool the microprocessor-and probably cutting the cost of any laptop or Internet device it's in. Such improvements would help fuel the growth for mobile computing.

The Crusoe TM5400, aimed at laptops, will use power management software to adjust its own operating speed and voltage to use even less power, depending on the demands of the program a user is running. For example, a software program that needs only 350 MHz of power will get only that, even though the Crusoe TM5400 will be capable of speeds up to 700 MHz. Cutting processing speed by 10% reduces the power needs of a Crusoe-based laptop by 30%.

"What you could wind up with is a laptop chip that processes like a Pentium III but uses about 5% of the power," said Carolee Bigard, a technology researcher with Raintree Securities.

The "could" in that statement contains a couple of suppositions that have yet proved. One of the Web devices demonstrated at the Crusoe unveiling was a Web pad, a portable browser running on a particularly lightweight mobile version of the Linux operating system that will let users access full-sized Web pages on the run. The Web pad is the first Transmeta device to draw the endorsement of a computer manufacturer, S3 Inc., which hopes to market a model for less than $1000.

The bigger question is how willing computer manufacturers will be to build in chips from a manufacturer that doesn't even own its manufacturing capacity.

"It was significant that no OEMs showed up to that coming-out party," said Harriman. "I would expect that Transmeta will deal with small, second-tier suppliers first, while Intel readies its response. But you know IBM, Dell and Compaq will be watching."