The Extremely Unprofitable Transmeta
The marriage of x86 and VLIW
Dave Ditzel grew up in the American Midwest, a few minutes from Iowa State University. Both of his parents had university educations, and his father worked as a chemical engineer. His early interests in science and tinkering were encouraged by his parents, and he eventually started doing electronics repairs (mostly televisions) for neighbors. This eventually led to him having a decent collection of electronic components, and the young Ditzel became interested in amateur radio. By fifth grade, he had become interested in computer programming as well. He had a Teletype Model 33 and timeshare access, and the school computer had a BASIC interpreter. At the age of twelve, he built his own card punch and reader. By junior high, he had firmly established interests in both hardware and software, and this culminated in high school. Eight bit CPUs were commercially available, so he made his own logic board, memory board, storage system with cassettes, paper tape, and a keyboard. For this endeavor, BYTE magazine was his primary instruction manual. While tinkering on his own, he also picked up summer jobs at the University working on the SYMBOL computer system (started by Fairchild, but Fairchild cancelled it, and the machine ended up at Iowa State). This is a job he kept when began attending Iowa State University. He also worked with System 360, PDP-11, and a few other contemporary machines.
Around the time Ditzel was completing his undergraduate education, some recruiters from Bell Labs were on campus. He mentioned that he had an interest in building a computer for C programming, and they responded well. At that time, Bell had a group at Murray Hill in New Jersey working on that very same idea. Ditzel got the interview, and Bell offered to pay for his graduate education. At work, he was on a team of about thirty people. For graduate school, he attended U.C. Berkeley where he studied under the tutelage of Dave Patterson who had coined the term RISC and led the Berkeley RISC project.
After working at Bell for about a decade, Ditzel went to work for Sun Microsystems. His first tasks were around the 32bit SPARC ISA and recruiting Patterson to work at the company. A little later on, he worked on the 64bit extensions to provide upward compatibility. Over time, Ditzel rose through the ranks at the company and became CTO.
Then, in 1990, some information was coming out about a new processor architecture in the USSR. Later that year, Bill Joy made a visit to learn about it, and this led to the creation of the Moscow Center for SPARC Technology. They were working on Very Long Instruction Word (VLIW) architecture that had begun in the Soviet Union with Mikhail Alexanderovich Kartsev and Boris Babayan. Dave Ditzel was among those working on Elbrus 3 with the MCST arm of Sun. A few years later, Sun canceled their VLIW project. Many people in the CPU industry were becoming disillusioned around this time as the wildness of the 1980s was quickly ending as Intel’s x86 together with Microsoft Windows conquered the world. Among this mass of the disillusioned was Ditzel. He longed for something exciting once again. He’d also begun to realize that binary translation would free chip design from software implementation. There was no reason to limit hardware for the sake of software. The cancellation of Sun’s VLIW project combined with the idea of binary translation gave him the push required to start a new company, Transmeta, in Santa Clara.
Ditzel brought in a few other key people when starting the company: Bob Cmelik, Colin Hunter, Ed Kelly, Doug Laird, Malcolm Wing, Greg Zyner. From the start on the 3rd of March in 1995, the company was incredibly secretive. NDAs were the rule of the day, and the company was so secretive that some of the faculty at Stanford referred to it as the cult, a perception not helped by the enthusiasm and excitement of the company’s employees. Part of this secrecy was likely due to where their first funding came from. While Ditzel aimed for a VLIW chip that would do dynamic binary translation from x86 to his new architecture, DARPA provided money for research into cryogenic electronics. While this early work was unrelated to the products that Transmeta eventually made, I cannot help but think that this early work may have led to the low-power consumption and great thermals that Transmeta later achieved. The provable reason for secrecy was an abundance of caution. Transmeta’s first patent was filed in 1996, and it wasn’t approved until 2000. Only after that did the company’s products launch. Anyway, after concluding their research work for DARPA, the company sought out venture capitalists among whom were Vulcan Ventures, Walden Venture Capital, Integral Capital Partners, and Novus Ventures.
The early culture of the company was seriously one of innovation. The ebullience shared by each employee gave the place the feel of a college campus except that each person had his/her own office. Patterson once referred to Ditzel as “the world’s best recruiter,” and recruit he did. Ditzel said, “we realized there were less than a dozen people in the world capable of doing this. We had one simple plan: hire all of them.” This is partially because he needed those people, but also, this kept those talented people from going elsewhere and competing with him. Two thirds of the employees were engineers, and they had been the top of their class at places like Stanford, MIT, Berkeley, and CalTech. The company offered quite a bit of lab space and equipment, and while it wasn’t required of them, workdays often turned into all night affairs. Of course, Transmeta didn’t limit hiring to recent college graduates. Having raised around $250 million to complete the product, they courted and won some experienced talent as well, and this only fueled speculation around the company. With Linus Torvalds (creator of Linux) having joined in 1997 and Dave Taylor (worked Doom, Quake, and their UNIX ports, Jaguar ports for Doom and Wolfenstein) in 1998, the hype was starting to hit a fever pitch. Within the company, there was a joke that their ticker symbol ought to be “HYPE” (of course, the symbol was TMTA when the company did IPO).
What Transmeta built was a four-wide, 128bit, VLIW processor that speculated… well, everything, and it did so quite aggressively. Instructions were referred to as molecules made up of either two or four atoms, where each atom was a 32bit RISC instruction. This allowed the chip to run two or four operations with a single instruction. The atoms of any molecule were then dispatched to five possible pieces of the CPU: two ALUs, a memory unit, a floating point unit, or a branch predictor. This CPU had 64 general purpose registers, and 32 floating point registers. The high register count (for the time) allowed for x86 registers to be assigned to native VLIW registers while also allowing their dynamic binary translation software to have its own registers. The idea behind VLIW is to have no real speculation and in-order operation. Traditionally, a VLIW chip takes a long instruction, breaks it into smaller instructions, runs them all in parallel, and provides the results. It does this quickly with (historically) four to eight instructions per clock cycle. The Transmeta design changes much of this.
This CPU was designed with the intent of using dynamic binary translation software, called Code Morphing Software (CMS), rather than targeting the underlying chip architecture. Within CMS, an x86 instruction would be decoded with particular attention to keeping memory access ordering while also checking the frequency of execution, branch directions, and I/O operations. Once enough executions are collected to fill out a molecule, the molecule is passed to the translator. Results of translation were cached allowing subsequent runs to be faster (at least until the cache is invalidated). Inherent to the design of CMS is rampant speculative execution (work being done before the need of that work is verified), and both hardware and CMS had methods to correct speculative failures. The primary method of CMS being to collect a failure rate and upon reaching an excessive failure rate to reinterpret completely in-order. In practice, this worked very well. On the hardware side of this, all x86 registers had both a working register and a shadow register. A failure in speculation would restore the register contents from the unmodified copy, and the series of instructions would then proceed without speculation. Both CMS and the translation cache reside in RAM (the CMS is copied from ROM during system initialization) which meant that on systems built with the chip and CMS, the RAM available to the OS and applications would effectively be reduced by 8MB or 16MB depending upon the particular implementation used. CMS also handled things normally done by a PC BIOS. One other bit of tech was what Transmeta called LongRun where CMS would estimate the power required by an application and adjust the CPU speed hundreds of times per second in increments of 33MHz (25mV). As CMS did its speculative activities, it could set the clock speed for the work to be done. Thermals, system load, and pipelines all contributed to LongRun speed adjustments.
While Transmeta’s first aim was compatibility with x86, they didn’t initially think it would end there. The Code Morphing Software was designed to be a general purpose platform to emulate any given architecture, run any given OS, and execute any given binary format. The hope was to take advantage of VLIW’s advantages while overcoming its disadvantages with CMS. With the hardware and CMS having been designed together, they felt confident that they could achieve the performance of Pentium III while using far less silicon, far fewer transistors, and far less power thereby producing far less heat. For those thinking that Transmeta would never be able to achieve parity due to translation overhead, it is worth noting that x86 CPUs translate every single instruction as well, and they do so with every execution.
The 1990s were a decade of transformation in the computing world, and x86 with Windows became the dominant force during that time. Yet, a whole new market had opened up over that decade. Mobile devices with internet connections were becoming more and more common. These devices needed a CPU that could sip power, and somewhere during the creation of Transmeta’s first CPU, the company made the decision to target that emerging market. This may have been due to the performance they observed with their first working silicon in January of 1998. While aiming for something competitive with a 400MHz Pentium II (they targeted the Pentium III not long before launch), they had something roughly on par with a 200MHz Pentium I. The chip also had some serious bugs like Tomb Raider 2 hanging after a quarter hour of play time, encryption not working in IE 5.5, MechWarrior II not playing properly, and modems not working in all PCI slots. Otherwise, it sipped power, was compatible with x86, and seemed to work in most cases. They set about improving things right away, and they quickly made progress.
Being conscious of the hype building around the company, they chose to fuel it on their website. Their website read: “This web page is not here yet” for several years. On the 16th of November in 1999 (I think), this message was changed to read:
Yes, there is a secret message, and this is it: Transmeta's policy has been to remain silent about its plans until it had something to demonstrate to the world. On January 19, 2000, Transmeta is going to announce and demonstrate what Crusoe processors can do. Simultaneously, all of the details will go up on this Web site for everyone on the Internet to see. Crusoe will be cool hardware and software for mobile applications. Crusoe will be unconventional, which is why we wanted to let you know in advance to come look at the entire Web site in January, so that you can get the full story and have access to all of the real details as soon as they are available.
As promised on the 19th of January in 2000, Transmeta announced their first CPU line, named Crusoe at an event at Villa Montalvo in Saratoga. Ditzel explained the hardware, the company’s intent, a bit about how and why their technology works, and what the company’s motivations were. To demonstrate the CPU and its x86 compatibility, Linus Torvalds and Dave Taylor played a 1v1 Quake match on the new chip. Torvalds lost... badly.
Crusoe came in two SKUs, the TM3120 (aimed at web appliances) and the TM5400 (aimed at notebooks). The 3120 could run at 333MHz or 400MHz (different SKUs), had a 96K L1 cache, came in a 474-pin BGA package, and was made on 220nm process by IBM on a 77mm die. It featured an integrated SDRAM controller and PCI interface. The 5400 could run at 500MHz or 700MHz (different SKUs), had a 128K L1 cache, a 256K L2 cache, came in a 479-pin BGA package, and was made on a 180nm process by IBM on a 73mm die. It featured an integrated DDR memory controller (in addition to SDRAM) and PCI interface. TDP for these chips ranged from 0.5W to 1.5W with an average of around 1W. The 3120 was eventually upgraded to run at 400MHz and sold as the 3200. The 5400 was later upgraded with doubled L2 cache and sold as the 5600.
Crusoe managed to achieve the goals that Transmeta had set for it. At 700MHz, the TM5400 matched the performance of a 500MHz Pentium III, or it did so most of the time. Crusoe’s southbridge had rather low bandwidth. This hurt performance considerably in any graphically or I/O intensive application. Still, if one were watching a DVD, for example, the 5400 ran at around 48°C without any cooling system. The average Pentium III with cooling could get up to its maximum operating temperature of 80°C to 105°C. Leaving heat aside, Crusoe was built with around a quarter of the transistors, and it used about a sixth of the power. Quite an impressive piece of kit.
Right away, one problem was obvious; the company had no customers at the announcement event. While a market for mobile devices consuming little power was growing (especially in Japan), it wasn’t quite there yet. In most portable devices, the CPU wasn’t even the primary power consumer. Such an honor would be given to the HDD, FDD, CDROM, graphics processor, or the screen. Add to this that customers weren’t yet thinking about things in terms of heat production or power consumption, and you have a product with no obvious buyer. Those who did see Transmeta’s announcement as a big deal were Intel and AMD. They’d already been in a clock speed race, and now they were going to aim their sights on power. Back to the topic of no customers having been present at the announcement, the first consumer device to ship with a Transmeta Crusoe CPU was the Sony VAIO PCG-C1VN PictureBook in September of 2000.
The PCG-C1VN was what we’d call an ultraportable or a netbook. The system weighed in 2.2 pounds and measured 248mm by 27mm by 152mm. It made use of a Transmeta 5600 clocked at 600MHz, 128MB of 100MHz SDRAM, an ATI Rage Mobility GPU with 8MB of VRAM (AGP), a 12GB HDD, a memory stick slot, 56K modem, 8.9” TFT capable of a resolution of 1024 by 480, an i.LINK S400 port (IEEE 1394), a type II PCMCIA slot, a USB port, stereo speakers, pointer nub with jog dial and three buttons, a microphone, a camera (1/6" CCD, 350,000 pixels, lens: f=2.8mm/F 2.8), and SVGA output (handled by an included USB to SVGA dongle). The standard lithium battery offered 2.5 hours to 5.5 hours of operation depending upon usage, but higher capacity batteries were available. A double capacity battery would yield 5 hours to 11 hours, and a quad capacity battery could offer 20 hours. This system shipped with Windows Me, and had several bundled applications: Acrobat, Quicken, McAfee VirusScan, Word 2000, QuickTime, Netscape Communicator, and RealPlayer. It also bundled several Sony specific applications: DVgate, Media Bar, MovieShaker, PictureGear Smart Capture, VAIO Support Agent. As was usual for machines of this era, several ISPs had their clients bundled: AOL, CompuServe 2000, EarthLink Network TotalAccess, Prodigy. External floppy disk drives and optical drives were available as well, and the Sony PC Card CD-ROM drive was about $300. Despite the Crusoe being less expensive than the Intel chips Sony had previously used in this line, this was still priced rather high at $2300 on account of its diminutive size and high feature count. So, in 2024 dollars, a consumer buying this netbook with CD-ROM (on which most software was shipped at the time) would need to shell out around $4760 before sales tax. PC Magazine reviewed the machine somewhat negatively, but Phil Hughes of the Linux Journal gave it quite a bit of praise. He carried the computer with him on a trip and estimated his battery life to have been between three and four hours. He noted the large feature count and he stated:
The VAIO is no toy. The excellent screen, fast speed and huge disk make it a real computer. I see it as the perfect system for either travel to remote locations where size and weight are really important, or cases when you need a part-time system, such as making a presentation at a tradeshow.
His primary complaint was the keyboard, and with that, his complaint was key placement as well as the size of specific keys. Otherwise, he noted that function keys lacked their special purpose uses in Linux due to these being handled by software on the Crusoe PictureBooks.
As for how Linux was running on the Sony device, unsurprisingly given the employment of Torvalds, Transmeta themselves created a distribution called Mobile Linux for their architecture (running with x86 compatibility via the CMS). In particular, they worked on things like touch screen support, on-screen keyboards, handwriting recognition, filesystem compression, and networking improvements.
Fujitsu and Sharp followed with products built around Crusoe. Compaq had signed on, but later backed out. In early November of 2000, IBM canceled their Crusoe-based ThinkPad, which had already been demonstrated to the public at PC Expo. About the cancellation, IBM made no comment. On the 7th of November in 2000, Transmeta had their initial public offering. The opening price was $21 per share, and they closed at $46 per share.
Dave Ditzel stepped down as CEO in 2001, and Mark Allen (previously at Cirrus Logic) took over for seven months before Murray Goldman (previously retired from Motorola) took the reigns in October. Goldman had become Chairman back in November of 1998.
A rather significant problem facing the company was manufacturing delays. As Transmeta was fabless they relied on IBM, and IBM was having some problems. The following year, IBM laid off 1500 personnel from IBM Microelectronics (their fab division). The IBM foundry business continued to decline thereafter and was eventually sold to Global Foundries… but IBM paid Global Foundries to take it. In response, Transmeta established a relationship with TSMC. This required a small modification to their design, but it also brought some improvements. The 5500 and 5800 are the equivalents to the 5400 and 5600, but these were built on the TSMC 130nm process and could be clocked up to 1GHz. Embedded versions of these were made and rated to perform reliably for a decade. These versions were sold as the 55E and 58E. The final Crusoe revisions came as the 5700 and 5900. These dropped support for SDRAM and came in physically smaller packages. Still, moving from IBM’s 180nm process to TSMC’s new 130nm process took time, and that was really the one thing that Transmeta couldn’t afford; they were simply forced. From what I can tell, it would seem that both TSMC and Transmeta had issues. Crusoe was the first chip built on TSMC’s 130nm node, and Transmeta apparently had some design issues as well. Delays continued for some time. This, of course, hurt systems vendors who were counting on finished silicon to be available.
Matthew R. Perry (formerly at Cirrus Logic) became president and CEO on the 11th of April in 2002. He also gained a board seat, but Goldman continued as chairman. On this change in company leadership, Goldman stated:
Matt has all of the skills needed to lead Transmeta, including demonstrated leadership ability, genuine technical expertise, sophistication in marketing and strategic matters, and a winning enthusiasm.
On the 18th of July in 2002, Transmeta laid off 200 employees representing 40% of the company’s workforce in response to a loss of $25.2 million for their second quarter. Transmeta’s stock had lost roughly 90% of its value per share. This was about the time that the 5800 began shipping in volume.
The Intel Pentium M (based upon the Pentium III, which in turn was based upon the Pentium II, which in turn was based upon the Pentium Pro) was released alongside the Centrino chipset in March of 2003. While the Pentium M used around three times more power and consequently generated more heat than Transmeta’s chips, it was also significantly more powerful, and Centrino was a solid chipset. It didn’t really help that Transmeta had nothing to show in response. On the 17th of July in 2003, Linus Torvalds left Transmeta to work on the Linux kernel full-time. Several of the founders had left by this time, and an exodus of recognized names in the industry didn’t ameliorate the rapid diminution of the company’s image.
Finally on the 14th of October in 2003, Transmeta announced Efficeon. This was a 256bit VLIW chip. Like its predecessor, it used CMS. This time around, however, Transmeta included support in silicon for MMX, SSE, and SSE2 (Crusoe had only supported MMX). Floating point registers were increased to 80bit, but the count of registers remained the same. Northbridge core logic was integrated on die, and Efficeon supported both conventional DDR and ECC. This chip aimed for compatibility with the Pentium 4, but it also featured a HyperTransport bus, NX bit, AGP 2.0, and had a 7 watt TDP at 1.5GHz, but as noted, Transmeta’s chips could dial down the clock as needed and conserve energy. Having a 256bit molecule, Efficeon could execute eight 32bit atoms per cycle. Efficeon also had more cache with a 128K instruction cache, 64K data cache, and 1MB L2 cache. The first generation of Efficeon, the 8600, was built on TSMC’s 130nm process, and these chips could reach clocks of 1.2GHz. The second generation of Efficeon, the 8800, was built on Fujitsu’s 90nm process, and these chips could reach clocks of 2GHz. The 90nm chip had a surprisingly small die at just 68mm compared to the Pentium 4 at 112mm. Packaged, Efficeon measured 841mm, and that’s a measure of a package that includes the northbridge on die (as opposed to Pentium M where that’s a separate chip). Despite being impressive, Intel had gained its place in low-power, mobile computing. Sales for Efficeon were never good.
Transmeta’s LongRun2 technology, announced in 2004, had improved upon the original. No longer was it just variable clocks based upon speculation, heat, and context switching, but LongRun2 reduced current leakage at the transistor level. NEC licensed LongRun2 from Transmeta in March of 2004. Later in December, Fujitsu did the same. In January of 2005, Sony became a licensee. This license to Sony became a major focus of the company. In March of 2005, Transmeta laid off 68 people reducing the headcount to just 208, and of those about a hundred were employed to work on LongRun2 for Sony. As a result, Transmeta’s first profitable quarter was Q3 of 2005 with a net income of $10.1 million about a decade after the company’s founding. Sadly (to me at least), the profits had nothing to do with their products, but rather with their IP licensing. Art Swift became CEO in 2005.
On the 11th of October in 2006, Transmeta filed a lawsuit against Intel Corporation in the US District Court of Delaware alleging patent infringement. In question were patents for adaptive power control, address translation, speculative address translation via two different methods, another two patents concerning RISC architecture with multiple typed register sets, and two patents for superscalar RISC instruction scheduling. Transmeta’s filing states that Intel violated ten patents total with at least the P6, Pentium M, Core, and Core 2 families of CPUs. Intel filed a countersuit in January of 2007 denying Transmeta’s claims and accused Transmeta of violating seven of Intel’s patents.
The company continued to lose money, and on the 7th of February in 2007, Transmeta’s engineering division was shut down. The company refocused on licensing IP. Transmeta had been 192 just prior to the engineering shutdown, but was reduced to 117. They closed their sales and support offices in Taiwan and Japan, and proceeded over the next half year to lay off more people, and Lester Crudele took the job of CEO.
On the 6th of July in 2007, AMD announced that they were investing $7.5 million in Transmeta. This temporarily bumped Transmeta’s stock to $0.94. The two companies had worked together some in the past with HyperTransport and AMD64, and AMD was now seeking for technological IP to flow to AMD. AMD gained access to LongRun and LongRun2, and they also licensed Efficeon which was then used in Microsoft’s FlexGo. This was a computer rent-to-own plan intended for the developing world, and also for use in cybercafes. The PCs in question used Transmeta Efficeon CPUs manufactured by AMD.
On the 24th of October in 2007, a settlement was reached between Transmeta and Intel. Based upon the terms of the settlement, it would seem that Intel was at fault. They paid Transmeta a total of $250 million with another $34 million directly paid to Transmeta’s executives. In exchange, Transmeta licensed several patents to Intel, transferred a small number to Intel outright, and agreed not to make an x86-compatible CPU ever again.
The final LongRun2 licensee was NVIDIA for a one time fee of $25 million on the 8th of August in 2008. Transmeta’s acquisition by Novafora was announced on the 17th of November and completed on the 28th of January in 2009. Transmeta’s 140 patents (with more pending) were acquired by Intellectual Venture Funding on the same day the acquisition closed, and that sale was completed on the 4th of February in 2009. Novafora collapsed in July.
Transmeta is a story of exceptionally bad timing. The company launched its first CPUs during a massive recession. While the chips were less expensive, laptops and other mobile computing devices were not. Likewise, the hyperconnected world into which Transmeta had hoped to lead the industry wasn’t yet realized. By the time that Transmeta had readied their second attempt, Intel had already moved, and unlike Transmeta, Intel had its own fabs and could guarantee product delivery. I personally love the ideas that went into Crusoe’s design, and I wish that Efficeon had been better able to compete. VLIW is darned interesting. While Transmeta was a spectacular failure outside of Japan, it succeeded in one important manner. Crusoe forced Intel and AMD to consider power efficiency and thermals, and this brought those aspects of computing to the minds of the market more broadly. For nerds like myself, Transmeta proved that VLIW could be viable. On this last point though, I am certain that many Russians would argue that Elbrus, from MCST, did that and continues to do so (complete with x86 binary translation).
I have readers from many of the companies whose history I cover, and many of you were present for time periods I cover. A few of you are mentioned by name in my articles. All corrections to the record are welcome; feel free to leave a comment.