The Quick Transmeta Guide
In January 2000, after more than 4 years of operating in "stealth mode", Transmeta finally announced its first Crusoe products. The little company had enormous courage and a product that amazed the tech press, but unfortunately their products never gained traction in the marketplace. In 2005, they unofficially exited the microprocessor business and became dedicated to licensing their various technologies to the remaining players. In 2007, they officially became a licensing-only company. And in 2008, the company was sold to fabless semiconductor company Novafora.
Crusoe, and the later Efficeon, was the realization of a philosophy very much at odds with the rest of the market. The design of Crusoe was built on the idea that microprocessor cores had become much, much too complex. By this point in time x86, a Complex Instruction Set Computer (CISC) architecture, had taken over the entire consumer market. The Reduced Instruction Set Computer (RISC) philosophy, which advocated using an execution core capable of relatively few, simple operations, seemed to have died off; even putative RISC designs were often mind-bogglingly complicated.
Transmeta struck out in a very different direction. They built a processor that genuinely did have a very simple execution core, and then wrapped that core in some very powerful software tools that allowed the chip to do all the fancy things that its competitors could do. In fact, it used this software layer to emulate an x86 processor, thus allowing it to run x86 software without a hitch.
Transmeta's products used a potent "Code Morphing" technique to translate x86 instructions into internal Very Long Instruction Word (VLIW) instructions (called "molecules"). This is similar to, but more powerful than, the instruction translation that AMD and Intel had used since their K5 and Pentium Pro products. The Code Morphing engine would in effect recompile x86 instructions to VLIW on the fly. It could do some impressive tricks; it would cache the results of previous translations (so that they did not have to be retranslated) and keep track of how many times it had performed the same translation (so that it could focus its efforts on optimizing frequently-used translations). In other words, it got smarter the more it was used. Intel's Netburst microarchitecture would later borrow some of the same tricks.
Crusoe also had a very aggressive power management system, known as LongRun, that would vary both the clock frequency driving the processor and the voltage being supplied to it. Later, the Efficeon would introduce LongRun2, which took the same capabilities and added to them tweaks to the manufacturing process meant to produce more consistent yields. LongRun was capable of much finer frequency adjustments than the version of Intel's SpeedStep that was available at that time, and SpeedStep could not vary the processor's supply voltage. It also did this autonomously, with no modifications to the operating system or applications required.
All of this, and much more, was done entirely in software. While products from AMD and Intel did their register renaming, speculative execution, and instruction translation with specialized hardware, Transmeta's products used specialized software to perform the same operations with the execution core's simple general-purpose hardware. In fact, they did this along-side the execution of code associated with the operating system and applications.
This was the blessing and the curse of Transmeta's approach. The software-centric design made revisions much easier (in concept, a processor could be loaded with new code in suto) and meant that the same hardware could potentially be adapted to execute code targeted at the Alpha, PowerPC, or virtually any other architecture. And the simplicity of the associated hardware meant that it drew as little as one sixth the power of its competitors in the same tasks. But there were negatives as well. Because the same hardware had to handle both internal bookkeeping and user applications, the two types of code could be forced to fight for limited resources. And the software needed to make all of the whiz-bang features work was, in fact, quite complex.
Crusoe had been in development for years before the big public announcement. Transmeta had become the subject of much speculation in that time, as they actively recruited very smart people to work on their very secretive product. When they finally dropped the veil of secrecy, the trade press was thoroughly impressed.
Yet all was not sunshine and roses. Transmeta's primary target market was the nascent laptop market. None of the established players had a fortified position in that market yet, and Transmeta's good performance/low power design was a natural fit. Unfortunately, Intel was eyeing the same market. Competing with Intel's marketing prowess proved quite a challenge.
It got worse. Not long after Crusoe's launch, AMD and Intel began a much-publicized clock speed war, each working full steam to introduce ever-faster products. A key component in their ability to do this was the integration of their design and manufacturing processes; both companies owned their own fabrication plants, and could carefully tune their designs and their fabrication processes to squeeze out every last bit of performance. But Transmeta was fabless; they relied on external partners (including IBM, Taiwan Semiconductor, and Fujitsu) to handle the actual fabrication of their products. They simply could not keep up in the clock speed race, and had to turn to other optimizations to try to keep improve performance.
Transmeta was also trapped by a catch-22 of the industry. Their lagging performance lowered sales, which denied them the economies of scale of larger production runs. This in turn kept prices higher, which further damaged sales. Try as they might, they could not break out of this situation. The 2003 launch of the Efficeon, essentially a greatly enhanced Crusoe, failed to make the gains the company had hoped for.
In 2005, Transmeta sold off its processor division. This was widely seen as the effective end of the company as a player in the microprocessor world. They limped on until 2007, when they officially bowed out of the processor business and announced that they would transition to relying solely on licensing as their source of revenue. It was at that point that they laid off a large portion of their remaining staff and essentially admitted defeat. In September 2008 the company was put up for sale, and in November 2008 a deal was announced with fabless semiconductor company Novafora.
Editor's note: This article is part of the expansion plan for the Big Processor Guide. It will make a lot more sense in the near future.