One of Silicon Valley's most secretive companies, Transmeta, has finally unveiled a chip that could revolutionise computer design. The Intel-compatible Crusoe, named after Daniel Defoe's castaway, uses innovative techniques to reduce power consumption. The chip can run more than 20 operating systems, including Microsoft Windows, and should give portable machines much better battery life. Crusoe consumes less power because Transmeta has moved many of its operations off the silicon into software. This degrades performance, but in the mobile applications for which Crusoe has been designed, longer battery life may be more useful.
However, it must be admitted that the worldwide coverage of Transmeta's launch had little to do with its patented approach to emulation software, called "code-morphing", or its use of "very long instruction word" (VLIW) chip design. Most of it was attracted by the presence of Linus Torvalds, whose Linux software kernel is used in the increasingly popular GNU/Linux clone of the Unix operating system. So far, Transmeta has announced two Crusoe chips. The TM3120 is aimed at portables and internet appliances costing about $500 to $999, while the faster TM5400 is aimed at $1,200 to $2,500 notebook PCs running Windows. The chips are being made by IBM, and prices range from $65 for a 333MHz TM3120 to $329 for a 700MHz TM5400.
Crusoe uses Transmeta's impressive LongRun power-management technology, which is similar to Intel's SpeedStep. The processor can run more slowly to consume less power. Transmeta's chips use only 1 watt where rivals may use 4-10 watts, and there's a special "deep sleep" mode that uses only 20 milliwatts. Transmeta has also put a lot of work into developing Linux - the Torvalds connection - for use in "webpads" and internet appliances. No operating system dominates this nascent market, but there are numerous competitors. These include Be's BeOS, GEOS, QNX and Microsoft's Windows CE.
The brains behind Transmeta is chief executive officer David Ditzel, who founded it in Santa Clara, California, in 1995. He started the Risc (reduced instruction set computer) chip industry in 1980, when he co-wrote a paper, The Case For Risc, with David Patterson, a computer science professor at the university of California at Berkeley.
Later he joined Sun Microsystems and helped to create its Sparc line of Risc chips. Computer processors were getting bigger and more complicated as new instructions were added to provide the capabilities programmers wanted. Patterson and Ditzel suggested using fewer but more regular instructions: this would make it possible to make smaller, cheaper, faster processors. Programmers would not go insane writing software for these Risc chips because the code would be written by special compilers. These programs would convert the software writers' complex commands into low-level Risc code.
The Risc idea was hugely successful: Hewlett-Packard, Acorn, Mips, IBM, Intel, Motorola and many others adopted it. Mips chips power millions of games consoles, including the Sony PlayStation. The Acorn Risc Machine became the Arm processor, used in many successful products. The multi-chip IBM Risc architecture developed into the PowerPC used in Apple's machines. However, over the past few years, Risc chips have become larger and some are more complicated than the complex chips they were meant to replace.
Today's complex chips have also adopted Risc techniques. For example, Pentium chips convert x86 instructions - eg commands from Windows applications - into low-level instructions that can be executed quickly by a Risc core. This is where Crusoe comes in. Why not do that conversion in software, outside the chip?
Transmeta's code-morphing has several advantages. The chip itself can be made smaller and cheaper. The software part is more flexible. It can have bugs fixed on the fly, and could even be changed to provide compatibility with different processors, not just the Pentium/x86. The Crusoe software also optimises its "translations" (recompilations) for the Crusoe hardware and stores the results for reuse. This compensates for the initial loss of performance.
Crusoe uses "very long instruction word" (VLIW) processing, which has often been seen as the next big advance after Risc. VLIW was pioneered by Multiflow (www.multiflow.com), which was founded in 1985, and successfully emulated a Sparc chip in 1989. Hewlett-Packard and Intel used VLIW ideas in the next-generation Intel chip, Itanium (code-named Merced).
The idea behind Risc was to have a steady stream of simple, regular (fixed length) instructions for the chip to process. The idea behind VLIW was to package a number of these simple instructions together and send them to the chip in one operation. In Transmeta"s case, the software delivers four 32-bit operations in each 128-bit "word". (Transmeta calls the operations "atoms" and the words "molecules" but the idea is common to VLIW processing.)
Getting it to work efficiently isn't easy. If the software has only one instruction for the chip to process, it has to pad the "word" out with three "no-ops", so the result will be slow. If it can send three or four instructions every time, it should be really fast. At the moment, we don't know which will apply.
Getting Crusoe chips to work at all has been an achievement, but now they have to be used to create products people are willing to pay for. That's the hard bit.
Transmeta: www.transmeta.com. ZD TV did a live webcast of the Transmeta press conference. You can watch it at www.zdnet.com.
