Google
 

 

 

Transmeta

Intel Reaches for the Maalox

Francis Vale

Judge Thomas Penfield Jackson got it right, but he was way too late. It no longer matters whether the Supreme Court or the US Appeals Court votes his decision down in favor of Microsoft or not. For proof, just have a look at Transmeta. Here is a company that had been working in secret for almost five years to revolutionize microprocessor design. And what happened? They came out in January 2000 with an incredibly daring new suite of technologies that in the end will only serve to further cement the Windows hegemony. One of their radical new microprocessors has even been specially optimized for low power consumption in portable Windows systems. When you have assembled one of the best and brightest groups of software and hardware engineers on the planet and all they can do is come up with a way to make Windows 98 work better, you know it's all over. Even more incredible, Linus Torvalds, the uber meister of all things LINUX, played a key role in this irony drenched story that ends with making Mr. Bill even richer. Can this tale get any stranger? How did it come to this? As Deep Throat said to Woodward & Bernstein, you have to follow the money to find the answer.

It all began with RISC. When the RISC idea first stormed on the scene back in the early 1980's it was envisioned as a microprocessor revolution. RISC was a call to arms against the increasingly baroque, costly to design and produce, CISC (Complex Instruction Set Computer) architecture. The move to RISC was spurred in major part by Transmeta co-founder David Ditzel who was also the co-author of a seminal paper (with D.A. Patterson) in 1980 that argued for RISC. This paper, "The Case for the Reduced Instruction Set Computer," led to the creation of numerous new CPU adventures. There was MIPS, the Motorola 88000, the Digital Equipment Alpha, SPARC, the PowerPC from IBM and Motorola, and several others.

Along this anti-CISC war path, several (in)famous industry consortia appeared, such as the Apache group, that rallied around the idea of producing RISC-based Windows PCs. RISC's goals were thus significantly expanded. They now included dethroning Intel, the progenitor of the incredibly successful x86 CISC design and reigning PC CPU king. But in the end, only one RISC chip design, the Alpha, would truly emerge as an Intel-alternative Windows (NT) platform, and even that has just about faded away. What no one took into account in all those Sand Hill VC business plans was that it was a huge pain in the butt for developers to continually recompile (and usually redesign) their code to run on non-x86 RISC CPUs. Given the lowest common denominator market effect, it was a brain dead business call to simply say the hell with taking this RISC and sticking with mother market WinTel. Today, the only arguably successful RISC desk top survivor is the PowerPC, now powering not Windows, but a Mac near you. Finally, the highly elegant RISC design originally envisioned by Ditzel, who had since moved on to become the director of SPARC Labs and the CTO of Sun Microsystems, was itself becoming increasingly complex, bigger, hotter, and ever more power hungry. The baroque had come full circle.

RISC disenchantment was thus widespread come the 1990's and Silicon Valley was by now littered with the VC sun bleached bones of alternative architecture PC companies. Moreover, during the 80's and 90's the "we-don't have-a-monopoly" MS Windows/x86 duo had gone on to capture almost 90% of the worldwide desktop market. So if you are RISC godfather Ditzel in 1995, it's obviously time to sit down and rethink your they-gotta-go RISC/CISC position. His solution was ingenious, and breathtakingly daring. Essentially, he said let's do away with the whole notion of hardware-bound CPU architectures, CISC or RISC, and let's make a superfast, lean and mean chip surrounded by an equally intelligent software shell. Software will rule, not logic gates. This software shell would take any CPU-specific operating system and its applications and transform them all on the fly to run on a fast running - but power miserly - small hardware core. In effect, this novel soft shell CPU crab would be, as Transmeta puts it, an "emulator on steroids."

If you could blazingly fast emulate any operating system/application/CPU combination on the planet, which one would you choose? Duh. Hello, WinTel, the only business call you could make if you are the mid-90's Ditzel and want to get funded to bring this crazy new notion to fruition. But if you are out pitching fast, cheap, cool running, small factor CPUs, what about also going after the mobile device market, like PDAs, smart cell phones and such? Again, show me the money rules. This mobile market segment was, and still is, already well served by any number of formidable CPU competitors with equally cheaper, smaller, cooler stories to tell, like the StrongARM. Whether it was brilliant foresight or dumb luck hindsight, Transmeta made the right business call at the time. Because if you can make very lightweight, mobile x86 clones that offer the "rich Windows experience," plus have the ability to run off just one battery charge two full length DVD flicks and still have four or five hours left over to do your 10,000 column Excel expense report on that long plane ride to Beijing, you have a great marketing pitch to start the new Millennium.

Thus, Ditzel got his startup money to build radical new microprocessors whose eventual mission was to make Windows look good (at Transmeta startup in mid-1995, LINUX was still a distant dark horse). To date, according to Bill Tai, a Transmeta director and its first VC investor, Transmeta has raised "in the hundreds of millions of dollars," with the first VC round led by Mr Tai's Institutional Venture Partners and the Walden Group (where Mr. Tai was a former partner); a second round was led by Paul Allen's Vulcan Ventures, a third round by IVP and Integral Capital Partners, and a forth round by George Soros with particpation from Deutsche Bank. Recently an additional round led was led by Sony, Compaq and Gateway which closed several months before the IPO filing in mid-August 2000. Indeed, the IPO prospectus shows $40 million in historical revenue during the Company's top secret start up phase, in addition to over $112 Million in cash on the balance sheet as the company heads into its IPO. Earlier, Tai also let slip that Transmeta has received help from IBM, which provided people, time, wafers, and other resources. In Tai's opinion, the subjective value of IBM's non cash contribution could be worth "under $100 million but more than $10 million."

But what Mr. Tai did not say about IBM was that Big Blue also provided something much more valuable to Transmeta than soft dollars: it also gave Transmeta legal cover from the notoriously litigious Intel. IBM has a cross-licensing agreement with Intel, and all those Intel-agida Transmeta chips being made at IBM's Essex Junction, Vermont fab plant come under the protective cover of that agreement. Without that lawsuit missile defense system, Transmeta could easily have seen all its money in the bank - and its potential customers - go up in a big legal mushroom cloud.

And so in January 2000, Transmeta lifted the veil of secrecy that it had managed to keep in place for four and one half years (an unheard of accomplishment in Silicon Valley) to reveal to the world two new, fully x86 compatible, Transmeta processors, the TM3200 and the TM5400. The TM3200 is aimed at Internet appliances and ultra-light mobile PCs, while the TM5400 is targeted at high-performance, "full-featured" 3-4 lb. mobile Windows PCs. Remarkably, just as Ditzel originally envisioned, all of the x86 architecture was emulated in software, right down to the registers. More to the marketing point, the unveiling ceremony showed that the Transmeta powered prototypes on display could run Windows well, fast, cool and long.

In addition to Windows, the Transmeta chips also ran Mobile LINUX, a new strain of standard x86 LINUX that will eventually be Open Source. This new, on-the-move, ultimately all flash ROM-based code had been shepherded into existence by another Transmeta luminary, Linus Torvalds, a man whose name is synonymous with LINUX. Torvalds had joined in Transmeta several years back and his ever growing stature added much to the mystery buzz that surrounded the company until the 2000 unveiling. It has been reported that Torvalds did much more than just develop Mobile LINUX. His legendary coding and debugging skills also played a significant role in the development of the highly innovative Code Morphing Software (CMS) technology that powers the Transmeta chips, which go by the brand name "Crusoe" (as in the peripatetic Robinson).

Code Morphing is just that. It morphs x86 binary instruction calls made by the operating system and transforms them on the fly into code that executes on the Crusoe machine inner core. Nothing has to be modified in the x86 operating system code to run on a Crusoe platform. To a user, it looks just like any other PC compatible system. And just as critically, the software developer never has to recompile to make an x86 Windows/LINUX/Solaris86/BeOS/etc. application work on a Transmeta powered machine. The Code Morphing Software is the only thing the x86 binary code ever sees, and completely insulates an operating system, including the PC BIOS, from the Crusoe's native instruction set. This no-compile-no-worry strategy eliminates one of the biggest market stumbling blocks too late discovered on the high RISC rocky road. Because all the complex application/operating system/x86 morphing action takes place within the CMS outer shell, the Transmeta processor at the system core is therefore small, simple, efficient, and cheap to make. The Transmeta chip only requires one quarter of the number of logic transistors found in an all-hardware design of similar processor performance. This processor, because it is so simple, also does not require huge design teams to design, develop and debug, as is the case with either RISC or CISC architectures today.

This simplicity also means you can crank out new Transmeta Crusoe processor designs much more quickly, and via CMS, not worry about backwards compatibility with legacy software, a huge cost and complexity consideration for the x86 family. Transmeta can now adjust and emphasize different processor performance factors without worrying about disrupting the x86 software application/system market base. They just redesign the CMS shell to accommodate the new hardware core, with CMS still keeping the operating system and its applications blissfully ignorant of all the underlying changes.

This Code Morphing software, says Transmeta, is fundamentally a dynamic translation system, a program that compiles instructions for one instruction set architecture (ISA), in this case the x86 target ISA, into instructions for another ISA (the Crusoe ISA). The Code Morphing software resides in ROM and is the first program to start executing when the Crusoe processor boots. (For better performance, the Code Morphing software copies itself from ROM to DRAM at initialization time.)Thus, the only program written directly for a Crusoe engine of any stripe is the Code Morphing software itself. Transmeta claims that even with all the intense morphing activity, its 700-MHz chip will perform at least as well as a PIII-500.

The success of this software-driven architectural approach was clearly on display at the Transmeta unveiling. The two Transmeta Crusoe chips, the TM3200 and the TM5400/5600 (the 5600 came shortly after the initial product announcements), are radically different designs, yet they both ran native x86 code equally well. The CMS shell, which had to be redesigned to work on the TM5400/5600, took care of all the TM3200 back-compatibility issues.

In one fell stroke, Transmeta has redefined the word "portable" software. We now have fully "transportable" applications, operating systems, drivers, et al. All of which begs the question, where does the operating system and the application demarcation layer reside if it's now all just emulation? Is MS Word now any less than MS Windows, which formerly did all the heavy CPU lifting? Prior to Transmeta, life was much simpler and Judge Jackson's decision to split the Redmond company in two was made in a much easier to comprehend computing universe. So who owns and controls the ultimate API now? Microsoft and Windows or Transmeta and its CMS? Clearly, the PC software control ball has the potential of moving into Transmeta's court, and that prospect is not going to go down too well with Gates and company, who must also be keeping a very wary idea on Linus Torvald, an on the record, avowed Microsoft basher. Indeed, at the January press conference, it was mentioned that making CMS open source was discussed almost every time Transmeta staff met around the watercooler. That would move the whole issue of ultimate API control out of one company and into the impossible to kill world at large, not a very comforting thought if your company headquarters are in Redmond, WA.

So just exactly what is this potentially industry-disrupting CMS and Transmeta processor? Fundamentally, it's a highly intelligent VLIW software compiler surrounding a VLIW instruction set hardware engine capable of executing up to four instructions in each clock cycle. In addition, design decisions were made to put some code morphing support right into the very simple VLIW engine itself, which has two integer units, a floating point unit, a memory (load/store) unit, and a branch unit. VLIW, Very Long Instruction Word, technology has been around for a long time. The first company to make a commercial attempt out of systems based on VLIW technology was the long gone Multiflow in the late 1980's. VLIW technology can be found today in a variety of systems, including the Philips TriMedia multimedia DSP processor. If regular software compiler design is an art, then doing a VLIW compiler right is black magic. It's not a task for the faint of heart.

VLIW is a parallel processing technology. It takes a string of instructions and groups them together for all at once execution. In the case of Transmeta, they call their entire VLIW 128-bit word "molecule", and the up to four individual instructions within that word are called "atoms." It's up to the Transmeta VLIW compiler to sort out these grouped instruction blocks and feed them in just the right way to the VLIW engine. If this special compiler gets it wrong, the VLIW engine will stall and performance suffers. To get it right, Transmeta uses proprietary predictive techniques that go beyond typical state of the art "out-of-order-engine" technology. Superscalar out-of-order x86 processors, such as the Pentium II and Pentium III processors, also have multiple functional units that can execute RISC-like operations in parallel. Conventional x86 superscalar processors fetch x86 binary instructions from memory and decode them into micro-operations, which are then reordered by out-of-order dispatch hardware and fed to the functional units for parallel execution. But this type of decoding and dispatching hardware is much more complex than the Transmeta processor√.s highly simple VLIW engine. All this special x86 hardware also requires large quantities of power-hungry logic transistors, which dissipate heat in rough proportion to their numbers. Charts supplied by Transmeta show a Pentium III processor playing a DVD with a lap searing 105¬° C (221¬° F), while a Crusoe processor model TM5400/5600 plays a DVD at just 48¬° C (118¬° F).

But there is much more to this performance story than just x86 out-or-order dispatch units vs. VLIW predictive algorithms. The Crusoe CMS system is also clever in the way it handles instruction translation. When an x86 instruction is executed, it is put in a special "translation" cache on the Crusoe processor. When an instruction request comes along, the Crusoe first looks in the translation cache. If the instruction is found in cache, the VLIW engine avoids another software transaction and uses the stored result at full processor speed. Based on today's benchmarks, the Crusoe chips will typically come out slower due to increased translation overhead. But the Crusoe, which learns over time how to optimize an application's execution speed, will get faster and faster as it goes along. However, all benchmarks, which were written for conventional microprocessors that always re-execute the entire instruction again and again, don't take this sped up learning curve into account.

So, this being the benchmark case, and if you were Intel, where would you zero in on processor performance in your marketing counteroffensive? Expect much non-CPU heat and very little light these coming months. And indeed, a set of much publicized benchmark done by PC magazine in the fall of 2000 showed the Crusoe part to be slower than its Intel counterpart. The fact that such benchmarks have little relevance or applicability in this Transmeta CPU instance is apparently beside the point. But what else is new?

CMS uses a variety of x86 code execution techniques, all optimized to further reduce system overhead and maximize processor performance, and all or some of which can easily confound traditional CPU benchmarks. In real world software, less than 10% of an entire code can continually consume more than 90% of the CPU cycles. To avoid this cycle-sucking trap, The Transmeta CMS utilizes a number of strategies, ranging from interpretation (which has no translation overhead at all, but executes x86 code more slowly),through translation using very simple-minded code generation, all the way to highly optimized code (which takes longest to generate, but which runs fastest once translated). A set of sophisticated heuristics help choose among these execution modes based on dynamic feedback information gathered during actual execution of the code.

Another performance enhancement unique to the Crusoe chips is their on-board integration of "northbridge" controller. In any PC, you have both a northbridge and a southbridge controller. The southbridge connects the PCI bus to peripherals like the mouse, keyboard, etc. The northbridge connects the PCI bus and the CPU to the main memory RAM, which they both share. The Crusoe on-chip implementation of a northbridge controller, an industry first, shortens the data path from RAM and in the process lowers electrical resistance that must be overcome as the data travels to the northbridge. In lowering the resistance, less heat is consumed, which directly translates into lower Crusoe power requirements than a conventional CPU.

Obviously, a tremendous amount of intellectual energy has been expended at Transmeta in their quest for low power, long battery life systems. This goes so far as to optimize a complete chip, the TM5400/5600, for longer battery life via what Transmeta calls "LongRun" power management technology. LongRun takes advantage of the fact that power varies linearly with processor clock speed and by the square of the voltage, so adjusting both can produce cubic reductions in power consumption. Unlike most conventional x86 CPUs used in mobile environments that regulate their power consumption by rapidly alternating between running at full speed and basically turning the processor off, the TM5400/5600 dynamically adjusts its power consumption by changing its clock frequency on the fly. The analogy Transmeta uses is like saving electricity with a bulb by either rapidly flicking the light switch on and off or by using a gradual, adjustable dimmer switch. LongRun is a CPU dimmer.

CMS, says Transmeta, also plays a role in lowering power consumption on the fly. It can continuously monitor the demands on the processor and dynamically pick just the right clock speed (and also adjust power consumption) needed to run the application. For example, while awaiting a keystroke in a Word document, the CMS will throttle back the clock frequency, and the voltage, while it will boost the frequency and power when a DVD software decoder is running, a very CPU intensive task.

The combination of LongRun and CMS make this a superior battery life extending system to Intel's SpeedStep technology, used in the recently launched mobile Pentium-III chips. SpeedStep has just two modes, low power and high performance, and it switches from one to the other based on whether AC power is available , unless the user chooses otherwise. LongRun dynamically chooses among as many as 16 voltage and speed levels, with no user intervention required.

Taken altogether, Transmeta says its chips consume only 1-2 watts, compared with 8 to 16 watts for other mobile x86 processors in the same performance range. In sum, the Transmeta chips mean that faster, lighter, mobile PCs with longer battery life, smaller batteries, and no fans or other cooling contraptions will soon be coming our way. For example, Sony and NEC Corp have Crusoe-based notebooks in production, and Hitachi LTD recently announced its line of Transmeta-based notebooks, which it says will deliver between 4.3 to 4.7 hours of battery life (depending on screen size) on its standard battery.

On the other hand, in June 2000, IBM announced that its ThinkPad notebooks would be using the Transmeta chips by year-end and claimed this new technology will extend battery life by 50%. But in November 2000, Big Blue pulled back, saying that it was not going forward with its ThinkPad 240 using a Crusoe chip. IBM√.s ways are sometimes mysterious, and this Blue Transmeta story is likely far from over.

LongRun and the TM5400/5600 are heavily optimized for use in portable Windows systems. Transmeta says that it noticed certain features in the Windows operating system, like Win98's Advanced Power Configuration Interface, and made changes to its Code Morphing software to improve lap top battery longevity. According to Transmeta, the LongRun power optimization is not limited to any one particular flavor of Windows. Transmeta also hastens to add that CMS works equally well for LINUX. No doubt, with Torvald's presence, they have an intimate knowledge of how to make CMS work well with penguin labeled machines. However, all the "Crusoe Processor Typical Power Consumption" figures listed in the TM5400/5600 spec sheet make note that "Typical DVD power is measured while running the Win DVD¨ 2000 player under Windows¨ 98SE" and that "Typical MP3 (playback) power (is) measured while running MMJukebox under Windows 98SE." Not a penguin in sight, anywhere in these numbers.

There are good reasons for that. First, DVD playback on a LINUX PC is an oxymoron. As of this writing, there is not a single, robust, commercial grade DVD software decoder available for LINUX. Second, Transmeta has made it clear that it is the TM3200, running Mobile LINUX, that is the target platform for those cute, slippery software birds in tuxedos. The TM5400/5600 is all about selling truckloads of devices to PC OEM's who want to market mobile Windows systems with a competitive edge.

And so Transmeta, a company built on the organizing principle of technology innovation, with the help of the primo rock star of LINUX, Linus Torvalds, appears hell bent on feathering the financial nest of none other than Mr. Gates himself. It's cuckoo, and like the cuckoo bird, follows a similar Microsoft pattern; you throw the rival hatchlings out of the nest and take it over for yourself.

And there is the potential here for an even greater irony. Microsoft, the master of the Cuckoo bird displacement strategy, is forever intent on killing any animal in the forest it considers to be even a remote rival. If it perceives CMS as a threat to its API world domination, Microsoft will have no compunction in allying with Intel to knock off Transmeta, no matter how good Windows runs on Crusoe chips.

As for Mobile LINUX putting a market dent in Windows, with companies like Gateway announcing that it will use Transmeta chips (presumably the TM3200) for the Internet access appliances it is building for AOL, it has a shot. But this is a very different market place than selling mainstream Windows lap tops. As was noted by Transmeta during its January press conference, most LINUX-based appliances have missed their marketing mark. A stripped down LINUX running on a non-x86 processor just does not give the user the complete and rich Internet experience. When such a user goes to open an e-mailed MS doc attachment or tries to access a browser plug-in (themselves complete x86 applications) and the appliance barfs, he or she gets very frustrated, very fast.

But making LINUX mobile, x86 application/Internet rich, and running in just 16MB of flash ROM, the Transmeta target, is not sufficient. Because if reducing frustration and making the user experience rich are the keys, how will the ordinary consumer react to a full blown LINUX environment? It's still for geeks only, at least until user friendly system shells debut, perhaps like the one upcoming from Eazel. But even assuming all that is done, running something like LINUX Star Office, a no cost office suite from Sun Micro with the most confusing, brain dead, designed by committee user interface yet designed, will only plunge a penguin initiate into the deep cold of MS Office/Windows despair. There is still much, much that needs to be done to make LINUX a rich, user friendly Windows alternative for the average consumer. To the great credit of the business people running Transmeta, they recognized this fiscal fact of computer life, and despite Torvald's integral presence, went ahead and built a Windows-optimized system. That Transmeta is not caught up in the death wish politics of an anti-MS jihad augurs well for its success.

On the other hand, we now painfully see that despite Judge Jackson's best attempts, nothing appears capable of stopping Microsoft, who always manages to somehow subvert even the most daring in new technology innovation. Like Deep Throat advised, you must follow the money, however unpleasant the trail.

Francis Vale, Copyright 2000, All Rights Reserved

21st, The VXM Network, http://www.vxm.com

s