Yes! The sacred day has finally come. The curtain of mystery over one of the most secret companies in the lastfive years is lifted. This company is Transmeta. And their offspring, we are going to speak about today, is a newprocessor aka Crusoe. As is known, the lack of credible information usually gives way to rumors, which have beenquite numerous recently. But now we can finally forget about all our guesses and vague ideas and find out the truth.

One of the first and most widely spread suppositions proved to be absolutely correct: Crusoe doesn't threatenthe desktop PCs from AMD and Intel. It could have been a real competitor to AMD and Intel products, if it had beenlaunched a bit earlier, but now… Anyway, it is not that bad. On the contrary, it consumes so little power that themanufacturers of portable devices, from notebooks to HPCs, may consider it just ideal for their products. However,we will return to this idea a bit later in our preview when we come to speak about some particular chip parts. Andto begin with, let's take a look at some more fundamental items. The first thing worth mentioning here is CodeMorphing technology, which allows transforming x86 code into the hardware engine's native instruction set on thefly.

By the way, since we touched upon the instruction set, we have to say that Crusoe belongs to VLIW processors. Inother words, unlike the ordinary chips working with CISC instructions, Crusoe uses VLIW (very long instruction word),which is a typical feature of such products as Merced and Elbrus 2000. (Especially, taking into account that the headof Transmeta Company, Dave Dietzel, spent a lot of time in Moscow and communicated with the future E2K developers.)

So, back to VLIW. This architecture isn't directly compatible with x86. However, nobody will ever disregard a greatlot of the already developed software for x86 processors. And Crusoe developers weren't an exception. They worked outan intermediate layer called Code Morphing. It consisted of a compact hardware engine surrounded by software layer.Code Morphing was transparent to programs and allowed transforming x86 instructions into VLIW on the run.

If we compare this approach with the traditional one, all the advantages and disadvantages will be just evident.They are:

1. Possible to change the processor structure drastically and to adjust it to these or those requirements,since all the changes of the kind can be made unseen for programs and hidden with the help of Code Morphing layer.
2. Possible to change the already manufactured processors again on the level of code transformation.
3. Good for different instruction sets, such as SSE, 3DNow!, etc. Get the license, and enjoy!
4. No matter what you do, but emulation remains emulation with all the possible consequences for the performance.To illustrate it we can offer you some harrowing results shown by x86 software emulation on PowerPC. However, Transmetadid a great job here, no doubt.

How does this all work?

To begin with, we should say a few words about the processor's logical structure. Crusoe core consists of 5 modules offour different types: 2 integer units, 1 floating point unit, 1 memory transfer unit and 1 branch unit. So, every VLIWinstruction, called a molecule, can be 64 bits or 128 bits long and contain up to four RISC-like instructions, called atoms.All atoms within a molecule are executed in parallel, each by a separate module. Molecules are executed in order, one by one,unlike most today's superscalar x86 processors with complex out-of-order hardware. This greatly simplifies the processorinterior allowing to give up some complicated functional modules (such as a decoder of numerous x86 instructions). Just tryto compare the die size of the mobile 0.18 micron Coppermine with the total cache of 288KB and the die size of TM5400 withthe total cache of 384KB: it will be 106 sq.mm against 73. These results directly tell on the level of heat dissipation aswell as on the power consumption.

However, let's not veer away from the main principles of processor functioning. On the layer of Code Morphing the initialinstructions get translated into molecules, and the latter, stuffed with atoms to their maximum, get to the next layerdescribed above from Code Morphing. Now we mean only x86, however in fact, nothing prevents the developers from designinga translator for a different instruction set. The entire processor environment starting from BIOS and finishing with theoperation system and software, communicates only with Code Morphing without any direct access to the processor core. It'svery convenient especially since both of the newly announced Transmeta processors have very different cores.

One of the methods used to increase the performance during this non-standard way of working is a highly logical cachingsystem. Each x86 instruction is translated once and is stored in a special cache located in the system memory. To carry itout once again you may leave out the translation stage and get the needed chain of molecules out of the cache. Besides,Transmeta also promised that Code Morphing would accumulate experience with the time: when the program is running, CodeMorphing optimizes it to make it work faster, pays special attention to the most frequently executed code segments, analysesthe branchings in the program body, etc.

The first Transmeta processors are mostly intended for the mobile chipsets market (which doesn't prevent the company fromdiscussing some sever versions with the same architecture). That's why the most meaningful parameter of this processor is itspower consumption. There are absolutely no problems with the youngest of the today's Transmeta chips - TM3120 with a 92KB cache.And as for TM5400, they built LongRun mechanism into it, which is functionally similar to Intel's technology called SpeedStep.This technology allows changing the processor clock frequency and core voltage on the fly. In fact, it appears much more flexiblethan in case of SpeedStep when we have two set frequency values.

Since we came to speak about some concrete chips, then let's devote it a bit more time. Today, there are already 2 first CPUsbased on Transmeta technology. The first one is called TM3120 and is designed for HPC market, and the second one - TM5400, seemsto be more suitable for subnotebooks market. So, let's begin!

TM3120 is the youngest processor in the family supporting the clock frequency of 333, 366 and 400MHz. Itpossesses only 96KB L1 cache (64KB for instructions and 32KB for data). The processor is designed for 1.5V.

It has some elements standard for PC-on-a-chip: 66-133MHz SDRAM controller with 3.3V interface intended for standardmobile applications SO-DIMM modules. The memory frequency is obtained with a certain divider (from 1/2 to 1/5) applied tothe processor clock frequency. The PCI bus controller also has the usual 33MHz PCI 2.1 compatible interface, which ensuresefficient work with all today's applications, which can make use of it. Speaking about the peripherals support, we shoulddefinitely mention a flash-memory controller: since this product is aimed at the market of the tiniest PCs, it simply can'tdo without this memory there.

As any other normal processor intended for Windows (since the list of reliable operation systems includes MicrosoftWindows 95, Windows 98, Windows NT and Linux), TM3120 has to comply with ACPI power management. And it does comply withit, consuming only 0.015W in Deep Sleep system state. Of course, if the processor carries out some complex calculations thepower consumption gets several times higher. For example, it grows up to 2.9W during DVD playback, which is also damned goodif compared to mobile Intel or AMD processors.

As for the second version, TM5400, everything mentioned above can also refer to it with a few corrections, though. First,the working frequency increased up to 500-700MHz. Second, which is also important, the cache size significantly increased:besides, 124KB (64+64) L1 there also appeared a 256KB L2 cache. Here the core voltage depends on the frequency and variesfrom 1.2 to 1.6V. However, the power consumption grows only up to 1.8W even during DVD playback due to the effect made byLongRun.

As for the PC-on-a-chip features, there is only one addition here. However, this addition is so significant, that itsimply can't remain unnoticed. The existing SDRAM controller got a company: a DDR SDRAM one. Everything we have said aboutthe SDRAM speed also refers to DDR.

The first chip is already being manufactured by IBM with 0.22 micron technology and costs somewhere between $65 and $89depending on the clock frequency. The second one should be manufactured in the middle of the year. It will be made with 0.18micron technology and will cost $119-329.

Well, all this is just brilliant. But what about Crusoe in real life? In fact, it proved not that bad for the start. Theyhave already successfully run the mentioned operation systems on a PC assembled on Crusoe CPU. Such real applications as PowerPoint, Quake, etc. Worked just perfectly.

As for the performance, Transmeta acted quite cleverly. They offered their own test - Mobile Platform Benchmark. We have toadmit that this benchmark is very logically arranged since it regards the performance-to-power consumption ratio as of paramountimportance. (And as for Crusoe's success here, we think it is the competitors who should care about it). According to the resultsachieved by Transmeta, TM5400 working at 266-533MHz (LongRun on) in some managed to keep pace with mobile Pentium III 500 teststhough in other ones it still lagged behind. However, it consumed several times less power than its competitors in all the testsrun. Here we can also mention one more result, though again according to Transmeta: 667MHz TM5400 performs almost at the samelevel as 500MHz Pentium III.

Well, it's high time we drew our conclusions. Judging by the announced parameters, we can say that this processor was worthwaiting for and the technology is very interesting at least from a theoretical point of view. Will it be of any particularinterest to large notebook manufacturers? Hardly, at least in the beginning. Besides, Transmeta was a bit late with the clockfrequency it provided. In the middle of the year mobile Intel processors are expected to surpass Crusoe here and as a ruleeverybody is hunting for megahertz. All in all, very much depends on the company's ability to convince the notebookmanufacturers of its reliability and on Crusoe's hidden potential. In other words, we will very soon make sure ifTransmeta turns out a good stayer or if it gets absolutely exhausted after another 100 meters…