This Santa Clara, CA company and its new Crusoe chip are attracting some of the best and the brightest minds in computing. So what exactly has Transmeta been up to, and how did they get Linus Torvalds to work for them?
But in January of this year, the veil was lifted, and the true story of Transmeta and what exactly they’re doing with Torvalds could be told.
It turns out that a major chapter in the story occurred in Seattle, where, three years ago, Transmeta CEO David Ditzel found himself pitching his vision for a whole new kind of computer chip to one of the founders of Microsoft, Corp., Paul Allen. It was 1997,and Ditzel was two years into a journey that could be traced back to a fundamental revelation he had while on leave from his job at Sun Microsystems. Ditzel saw the need for a clean slate in the development of computer chips.
In Allen’s small conference room perched above a 50-seat luxury theatre, David Ditzel was doing most of the talking, explaining how his chip would change the industry. He found avid listeners in the former Microsoft executive and his venture-capital scout Kevin Doren, who was also at the meeting.
Ditzel was nervous. Was Allen getting his message? It was not an easy one to grasp. About two-thirds of the way into the presentation, Allen began to smile and to laugh slightly to himself. All eyes turned to him. Allen recalled that at the start of the previous era — when Microsoft was born — he had written “a little simulator” himself. “I know exactly what you’re talking about here,” Allen said. “But what you guys are doing is a completely different way that we haven’t done before.”
The two men then really got into it. Ditzel went on about how removing the complexity from the chip and putting it into the software side instead would finally solve the problems of compatibility and intricacy that had hindered microprocessors from fueling small, swift, mobile computing devices that could connect to the Internet and provide functions previously unheard of. The idea of putting chip functions into software was not new, but Transmeta’s approach was different and promised high performance while vastly reducing the amount of electrical power required to run the device.
It all hinged on translation software that had never been written before,and — to be compatible with existing operating systems and applications — relied on a deep understanding of the x86PC- compatible architecture at the heart of most of the world’s computers. In essence, this software translation would allow the chip to mimic instructions used by other microprocessors. Transmeta also had to invent a new way of building the new chip for an entirely new architecture, never mind the computers that would actually use the processor.
Ditzel’s demonstration was winding down.
Paul Allen was on his feet and excused himself, followed by Doren. Ditzel and Hunter stayed in the conference room, packing up their demonstration, and believing it would be weeks until they’d get word on Allen’s response.
An hour later, they knew differently. Doren strolled back in the room.
Finally, he spoke up. “OK,” he said. “Paul wants to do it.”
Allen had thought the challenge would be “fun,” and he had his vision of what he called The Wired World, Ditzel recalls. “He had lots of different pieces of technology — graphics and display companies, and interesting software of various kinds — and the processor was the one thing missing from his portfolio.”
Ditzel was in a state of disbelief. “This is like the only venture-capitalist investor who could make up his mind in an hour!” he says.
Meanwhile, back in Santa Clara, at Transmeta’s office complex in a 10 by 10-foot room, Linus Torvalds was quietly working away on the company’s first iteration of its code-morphing software. It appeared to be an impossible challenge.
|Virtually Different: The Crusoe chip architecture.|
You could think of it as something like a Java virtual machine. Instead of write once, run anywhere, Transmeta’s philosophy is “write as many ways as you like, you can run it here.” Transmeta’s Crusoe chip contains code-morphing software that can trick any software running on the computer (the BIOS, operating system, and applications) into thinking that it’s talking with an x86 chip, even though Crusoe does not use the x86 architecture (it uses something called the Very Long Instruction Word, or VLIW). In fact, Crusoe could mimic any microprocessor architecture the Transmeta folks desired — Alpha, UltraSPARC, or MIPS — just by changing the code-morphing software (see diagram below).
Transmeta claims that there are two major benefits to this kind of architecture. For one thing, because software rather than the chip itself does the work of decoding x86 instructions, Crusoe needs far fewer transistors, which means it can work on smaller, more compact systems. For another, software is a heck of a lot easier to upgrade than hardware, so improvements to the Crusoe design can be passed on by simply changing the code-morphing software on your chip.
Changing the Rules
“I find it’s inspiring to go out and watch over the whole valley and wtach what people are doing. And try to figure out how to be one or two chess moves ahead of them.” – David Ditzel
PHOTO COURTESY OF TRANSMETA CORP.
Ditzel’s efforts to simplify computer processors had started some 20 years earlier with what was known as Reduced Instruction Set Processors, or RISC. In 1980, when Ditzel was a graduate student, he and his adviser, computer scientist David Patterson of the University of California at Berkeley, published the first computer paper on RISC in an effort to bring microprocessor technology to a new level. Since that time, almost every microprocessor in existence has used RISC techniques. But battles over complexity and compatibility dampened the success of RISC. Both types of processors, RISC and x86 PC-compatible chips, over the years became increasingly complex in their designs. “Each year we hear about chips with more and more transistors and we are impressed with the big numbers about millions of transistors on a chip. Unfortunately this complexity has also brought the entire computer industry a lot of problems,” says Ditzel. Hence, chips were taking longer to design and debug, requiring the work of huge engineering teams. They were getting bigger and more expensive to build, and running hotter.
Another problem for RISC was that no RISC processor was ever able to get anywhere close to the same number of software applications and operating systems that the x86-based PC got, Ditzel says. “The dominance of the PC platform, particularly after those PCs got connected to the Internet, was just too much to overcome,” he observes. New software was always available first on the PC, and perhaps never on RISC chips.
These problems caused Ditzel to question the future directions of microprocessor design. One possible solution around the chip-complexity problem seemed to be a new type of architecture called VLIW, or Very Long Instruction Word. It was viewed to be simpler and faster. But VLIW required all new software and so would be starting in an even worse position than RISC processors.
Ditzel, who had served as chief technical officer for the SPARC chips at Sun Microsystems until March 1995, was disturbed by the situation. That month, he left Sun on sabbatical, expecting never to return. “I went hiking in the hills behind my house and tried to clear my mind to think about how one might design a different type of computer that could have both simplicity and compatibility,” Ditzel recalls.”But I knew that designing compatibility into a new silicon chip was inherently a very complex function.”
During his trail hikes above the Valley, Ditzel mused over the fact that traditional problems often didn’t have traditional answers. “There’s a famous saying, ‘If the only tool you have is a hammer, then everything looks like a nail,’” he says. The problem was that microprocessors were typically designed by big companies with huge silicon-fabrication facilities that might cost a billion dollars each. These “fabs” were their hammer, and therefore every solution to them would seem like it needed to be solved with more silicon — their “nail.”
That’s when he realized that the compatibility problem could be solved through software. He spent the next few weeks barely sleeping, working out a basic plan. Ultimately, he assembled a team of eight of the brightest computer scientists he could find. They became the founders of Transmeta and began working out of Ditzel’s living room.
Silicon Dream Team
“We knew we wanted to have a very good relationship with Microsoft. On the other hand we also have certain Linux connections. It was nice because that could help us keep a little bit of a balance between the two.” – Linus Torvalds
JEAN FRANCOIS PODEVIN
Everyone who talked to Transmeta had to sign a nondisclosure agreement promising to keep secret anything they might find out during the interview process. The hiring plan, he says, was like a miniature version of the Manhattan project, the effort to build the first atomic bomb. In that case, top physicists started disappearing from universities and no one knew that they were all going to a secret facility in Los Alamos, NM. Now top chip designers from Silicon Valley started disappearing from major computer makers and no one knew where they were going. “It took a while before people realized that the top designers were all going to the same place, Transmeta,” Ditzel says.
On the software side, recruiting was more tricky. There were very few people in the world who had skills in dynamic software translation. “Our plan was simple. Hire all of them,” Ditzel says. “Much of what we were doing was new invention, and this called for the best minds we could find.”
Soon word got out that something was going on at Transmeta, and top engineers began seeking out the company. Indeed, some of the most sought-after new candidates in the computer business come from Stanford University’s doctoral program in computer systems, Ditzel remembers.
One student was finishing his Ph.D. dissertation and decided that the best way to find out where he should work was to call a number of the references in his dissertation and ask their opinions. As he went down the list, tracking down each reference, he was shocked to discover that each reference had quit his or her job in the last year to go to some start-up company named Transmeta. But no one seemed to know what this Transmeta company did. He contacted Transmeta directly and said, “I don’t know what you are doing, but I sure would like to interview there.” He interviewed and decided that Transmeta was where he wanted to work.
Word spread at Stanford that “Transmeta was the cool place to work,” though no one really knew what it did. Transmeta was often referred to simply as “the cult,” a place where people go in and never leave.
Enter Linus Torvalds
Transmeta literally went to the ends of the earth to recruit the best talent on the planet. Prominently displayed on a wall inside the company’s Santa Clara headquarters is a map that pinpoints 20 countries around the globe that represent the various homelands of its employees.
The development of its code-morphing software is managed by an engineer originally from Spain. A German programmer is also in the group, and the lead architect for Transmeta’s VLIW hardware architecture is from Australia.
In the fall of 1996, Peter Anvin, a Swedish programmer in the software group headed by Colin Hunter, approached his boss. “I know a really smart guy from the University in Finland,” he said. “He’s involved with Linux. In fact, it’s named after him.”
Why would Transmeta want an operating-system guy? Linus Torvalds was recruited by Transmeta not because of Linux, but because he had “incredibly good artistic taste at designing software,” Ditzel says.
He also was one of the few people in the world who really understood the x86 architecture at a very basic level from the standpoint of software engineering.
At the time, Torvalds was just finishing graduate school and was barely known. Linux was not yet the phenomenon it would become just a couple of years later. Ironically, Linux would become an important part of Transmeta’s strategy, although it represented only a tiny part of Torvalds’ job at the company.
Several software engineers at Transmeta had already been using Linux as a technical workstation. For a small startup, it was cheaper to buy x86 PCs running Linux than to purchase more expensive Unix-based RISC workstations for development. Programmers at the company who were familiar with Linux already had deep respect for the man who had written the kernel.
After a day of talks in Silicon Valley with Ditzel and members of the software team, Torvalds returned to his hotel room, shaking his head in disbelief. “Uhhh, these guys are crazy!!! Do they actually expect that to work??” he wondered. At the time, Transmeta had only a couple of emulators to show its concept. “The first version of the hardware-accurate simulator was much slower, and at that point was fairly flaky,” Torvalds says. But he thought the technology was exciting.
After sleeping on it, Torvalds returned to Transmeta’s office for another day of briefings. “I began to think maybe the reaction that these guys are crazy was actually the right reaction,” he says. “It meant Transmeta was doing something that nobody else was doing.”
In December, Torvalds’ first child was born, and a few weeks later the whole family moved to Santa Clara for Torvalds’ new job as a chief programmer on code-morphing software — one of many programmers at the startup.
“A lot of what Transmeta had done had been done in pieces at other places,” Torvalds says, “but what’s unique about Transmeta is that this is the first time anyone has put the pieces together and required a certain level of performance.”
While having Paul Allen as an investor “added to a lot of the mystique, because Paul had the interesting Microsoft connection,” having Linus Torvalds in Transmeta’s ranks balanced out the equation, Torvalds says. “We knew we wanted to have a very good relationship with Microsoft. On the other hand we also have certain Linux connections. It was nice because that could help us keep a little bit of a balance between the two.”
“The chip itself is like any other microprocessor — it’s a bunch of silicon, it has adders and caches and floating-point units, and the code-morphing software is just another program that runs on top of the chip,” says Torvalds. What’s different about what Transmeta has done is that the code-morphing software runs underneath the operating system, whereas almost any other application would run on top of the operating system. “So it actually tricks the operating system. The OS thinks it’s running on a bare piece of hardware. It’s the combination of hardware and software working together that gives the definition of what’s called a microprocessor. It’s always been just pure hardware in the past.”
What is Transmeta’s positioning for Windows systems versus Linux, seeing that it wants to play in both markets? “In many ways, Linux is going in certain kinds of areas and has certain capabilities where it is difficult for Windows to operate,” Ditzel says.
“What’s happening in Silicon Valley is people are building new widgets, new pieces of hardware, and making them do something not contemplated before. The only way they can make the thing work is to fiddle with the software, to add a new device driver or something else, and if they don’t have access to the source code it’s hard to do that sometimes, because Microsoft software right out of the box wasn’t set up to deal with handwriting recognition or voice recognition as a way to drive the device. So Linux provides that kind of opportunity to experiment, and I think once those things settle down, it’s perfectly complementary with the way Microsoft operates. They go out and wait for the market to grow, and once it’s been verified by a bunch of experimentalists they swoop in and do their thing. In many ways, we see ourselves fitting in in a complementary way. We have good relationships with both Microsoft and the Linux community.”
Indeed, Microsoft was briefed under NDA about Transmeta’s plans shortly after the company was founded.
It seems most companies briefed by Transmeta recognized the huge potential for the new type of hardware devices it would enable. International Data Corp. forecasts that in 2004, information-appliance sales will grow to $17.8 billion, or 89 million units, compared with $2.4 billion, or 11 million units, in 1999. IDC defines these as including palm-size and “clamshell” computers, Internet phones, thin clients, and set-top boxes — devices that analysts view to be well-served by the Crusoe chip.
So with a team of the best and the brightest of Silicon Valley, investment from Microsoft’s founder, and the creator of Linux on the payroll, Transmeta seems set on applying its new technology wherever the market wants to take it. Only time will tell how large a piece of that $18 billion pie it will have four years from now.
Wendy Goldman Rohm is the co-author of Under the Radar. She can be reached at email@example.com.