crusoe by stariya



      Scott Lasak

 Professor Robert Dewar


        Transmeta Corporation‟s Crusoe microprocessor has the potential to revise the

way microprocessors are designed and implemented in the future. The design moves

much of the processor‟s responsibilities from the hardware to software. This has

dramatically reduced the number of transistors compared to other current

microprocessors. In addition, the overall size of the chip has decreased. Coupled

together, these two attributes have produced a chip that performs like its counter parts,

yet produces very little heat and a low power consumption of about 1 watt.1 The Crusoe

can boast a power consumption savings of 60% to 70% less than other processors. The

special power management software, nicknamed LongRun allows the processor to

“dynamically adjust the processor frequency and voltage on the fly.”2 Want more? The

Crusoe still has one more truly remarkable aspect that Transmeta has not bragged about

yet. It may be that they are waiting for the right moment to make their mark on the

industry though. The Crusoe was developed to be fully compatible with the x86.

However, it is not limited to solely this compatibility. Instead, the software that takes

over so much of the hardware‟s responsibilities could be written to dynamically translate

any platform‟s code to the Crusoe‟s underlying hardware.

        To better understand the current “beast” that is Transmeta‟s Crusoe, a brief look

into its origins will reveal more about its design.     Going back to 1994, Apple had just

switched to the PowerPC CPU to keep up with Intel‟s 80x86 series. In order to maintain

compatibility with Apple‟s previous Motorola 680x0 series the new PowerPC needed an

emulator to run 680x0, one then turned out to run faster than the previous high end Apple

1 Transmeta’s Crusoe: A Low-Power x86-Compatible
Microprocessor built with Software.
  See footnote #1
Macintoshes. A new project was then started, after the IBM engineers took note of this,

to emulate the 80386+ architecture on a PowerPC. For whatever reasons that particular

project was later abandoned, but the concept was continued with the DAISY project

(Dynamically Architected Instruction Set from Yorktown). The DAISY, however,

translated code to a hypothetical VLIW (Very Long Instruction Word) CPU instead of

the PowerPC. The DAISY shared a particular feature with another later project called

Dynamo from Hewlett-Packard (which ran PA-RISC on PA-RISC). Both projects were

able to optimize code as they ran. In the case of the Dynamo, it could improve PA-RISC

performance by up to 20% over non-emulated code. From that point, several engineers

came together to found Transmeta Corporation. Transmeta provided the missing VLIW

processor that was previously only hypothetical and created a new dynamic translator to

emulate the 80x86, which they dubbed a “Code Morpher”.3

        Now lets take a look at a few overall characteristics of the Crusoe chip. It is

based on a 4-wide 128-bit Very Long Instruction Word (VLIW) implementation. On

board are actually five processing units: a load/store unit (LSU), a branch unit (BU), a

floating-point unit (FPU), and two integer units (ALU). It also includes sixty-four 32-bit

registers. Each VLIW instruction word can have either 64-bits/two instructions (32*2) or

128-bit/four instructions (32*4). By utilizing the VLIW design, a four-operation

instruction can utilize 4 different processing units to execute in parallel. Also, with two

integer units (ALU), an instruction word can even contain 2 integer operations. It does

not support variable-length instruction groups, which poses a compatibility problem that

is common among VLIW. However, only the translator runs directly on the chip, so

3 Transmeta Crusoe – Leaving Hardware.
(January 2000).
compatibility with other software is not really a consideration.4 It has an in-order 7-stage

integer pipeline and a 10-stage floating-point pipeline. The hardware generates the same

condition codes as conventional x86 processors and operates on the same 80-bit floating-

point numbers, in order to ease the translation process to the VLIW core instruction set.5

Because the Crusoe has a real VLIW under the hood benefiting from parallel instructions,

it has an instruction to select correct results after speculative execution. Emulated

registers are “check pointed” between blocks of optimized code so that exceptions cause

the block to be returned to its start point and the instructions can then be interpreted in

order. The Crusoe does this by taking the registers that contain information regarding the

x86 state and copies them into what is known as the “shadow copy.” If an exception

occurs, the working registers are dumped and the shadow copy is moved in. Loads are

protected and stores raise an exception, rather than automatically reloading from the

address. Multiple loads from a single address are moved into the exception handler, and

an alternate store instruction allows extra loads to be skipped if they are not needed. This

has the benefit of eliminating memory delays, instead of just reducing them as cache


         First lets examine the Code Morphing software that makes this processors so

flexible. The aim of the software is to dynamically translate x86 instructions into VLIW

instructions that are then interpreted by the core hardware. The scheme is laid out in a

kind of hierarchy, in which the hardware Crusoe engine is at the lowest level. Then when

the processor is fired up the Code Morphing software is the first to launch, then comes

  See footnote #3
5 Crusoe™ Processor Model TM5800.
  See footnote #3
the x86 BIOS which can then continue to layer upwards with the x86 Operating System

and then any traditional applications. The Code Morphing software actually consists of

two separate modules that work together: the Interpreter and the Translator. The

Interpreter module is responsible for interpreting the x86 instructions one byte at a time

and then dispatching execution to the underlying VLIW instructions. It also has the

added functionality of filtering infrequently executed code by gathering run time

statistical information about the x86 instructions. If specific x86 code has been executed

several times the second module, the Translator, is called upon. The x86 code is then

translated into highly optimized and extremely fast VLIW native instructions. These

native translations reduce the number of instructions, increasing performance.

Additionally, after execution they are then cached for future use, so that when the same

x86 code is required the highly optimized cached translations can be immediately

recalled without the need to retranslate. By utilizing a Translation Cache, Code

Morphing exploits the degree of repetition typically found in real world workloads.

While normally the cost of optimizing code on the fly can result in slower performance,

the initial costs are amortized over the carefully analyzed needs of re-executing and use

of caching.7

          “The Crusoe processor operates from a 0.9-1.3V core voltage supply at extremely

low power levels, even while the device is operating at very high performance.”8 With

conventional microprocessor designs approaching 40 million transistors, managing power

consumption can be challenging. Code Morphing already reduces the number of

transistors on the Crusoe by almost half that amount. In addition to their already energy

7 Code Morphing™ software.
    See footnote #5
efficient design, Transmeta‟s power management technology, called LongRun, further

reduces power consumption. LongRun grants on the fly power reduction by varying the

CPU clock speed, as well as, the core power supply voltage “in response to adaptive

power management protocols that monitor processor load demands and control processor

power and performance levels.”9 LongRun has the unique advantage of savings that can

be measured cubically, while Legacy Power Management technologies can only control

processor speed resulting in linear savings. Primitive techniques such as Clock

Throttling choke the processor performance by letting the processor run at full speed and

near ineffective performance. Even Intel‟s SpeedStep technology doesn‟t have the fine

amount of control that LongRun has, missing the opportunities for further power gains.

It‟s because LongRun has the ability to adjust both processor speed and voltage that it

can attain such low power consumption. The proof is that power varies linearly with

clock speed and by the square of voltage, adjusting both can produce cubic reductions

contrasting the linear savings of conventional CPUs. The LongRun technology works by

configuring the processor to be able to run at a number of different voltage and frequency

points. Then the LongRun algorithm embedded within the Code Morphing Software

monitors the Crusoe processor and is able to dynamically make changes between the pre-

configured points as runtime conditions change. The true scheme behind the power

savings is that LongRun is designed “to provide just enough performance for the

processor workload at hand.”10

        The compatibility of the Crusoe processor with x86 hardware and software is

 See footnote #5
10 LongRun® Power Management
perhaps Transmeta‟s greatest edge. The latest version, the TM5800, “was designed to

run with the full suite of x86 compatible operating systems, including Microsoft

Windows 98, Windows Me, Windows NT, Windows 2000 and the soon to be released

Windows XP. The TM5800 was also designed to work optimally with Linux and Midori

Linux Operating Systems.”11 While the company has definitely found its niche among

ultra-portable laptops and other mobile Internet computers (MICs), the ability to work

seamlessly with the markets most popular OS and software is a real selling point.

However, Transmeta has kept quite about the possibilities of working on other software

versions to it exceptional Code Morphing technology, and perhaps that‟s going to be their

greatest edge in the future. They might just have the benchmark to a completely new

approach to processor design. A single processor die, with low power consumption,

thermal management, and an extremely small size that could be mass manufactured with

different software modules that are obviously not limited to x86 implementations. It will

be a revolutionary proposal when Transmeta decides to “license its technology to or

begins manufacturing for traditional silicon vendors who see the opportunity to build off

a single die and do everything else in software.”12 Further, Transmeta has recently made

the move to the server market. With the proliferation of “server farms,” considerations of

space and electrical consumption are two important cost factors for companies. Many

companies utilize a Performance/Per Watt/Per Cubic Foot matrix to determine

profitability. Again Transmeta‟s Crusoe seems to have an edge to offer in another area of

computing. A Crusoe-powered server typically consumes at most one half the power of a

  See footnote #5
12 Dorsey, John. Transmeta Blends Hard and
Software on New Crusoe Chip. (January 19, 2000).
server designed with a conventional microprocessor, and due to its low power

characteristics, it can be designed without cooling mechanisms such as a fan or bulky

heat sinks. A higher concentration of servers within a given space and a given power

budget creates a much more appealing set of results for every company. A Crusoe-based

server is suppose to deliver a minimum of five times the performance of a comparable

conventional server using the same power budget.13

        Currently on the market, Transmeta Corporation offers five flavors of the Crusoe

processor. The TM3200 that clocks at 333, 366, and 400 Mhz and is targeted at Internet

mobile devices that draw on volatile memory instead of a disk drive. Coupled with the

Mobile Linux operating system (Midori) “users can expect a complete Internet

experience, including access to the full range of plug-in applications. Transmeta provides

Mobile Linux assistance to OEMs looking to accelerate their time to market with new

mobile Internet devices.”14 One additional note about Mobile Linux is that it can be

stored in solid state Flash ROM, removing the expense of hard drives required for other

larger operating systems. Then there are four other models in the Crusoe family:

TM5400 (500-700 MHz, 256K L2 Cache), TM5500 (0.13, 667-800MHz, 256K L2

Cache), TM5600 (500-700 MHz, 512K L2 Cache), and the TM5800 (0.13, 667-800

MHz, 512K L2 Cache).15 The new 0.13µm process technology allows the TM5800 to

run at speeds up to 800MHz while running significantly cooler than the TM5600. The

addition of the new CMS4.2 software and low power, high performance DDR SDRAM

13 Performance Under Server
Computing Workloads.
14 Transmeta Breaks the Silence,
Unveils Smart Processor to Revolutionize Mobile Internet Computing.
15 Transmeta Crusoe TM5x00.
memory the TM5800 can provide a boost in performance of up to 50%.16 Of course with

speeds of up to 800 MHz these models are targeted at the newer ultra-light mobile PC‟s.

The Crusoe, with its LongRun power management system, offers the longest running

mobile PCs at a high performance. PCs running the Windows operating system and

Microsoft Office applications could operate on a single battery for up to a full work day,

increasing the average user‟s productivity. Market blessing? “‟Our customers are telling

us that significant battery life improvement is the most requested feature by a margin of

two to one. That‟s why Crusoe‟s low power is so important,‟ said Transmeta‟s Jim

Chapman, vice president of sales and marketing. „The current mobile market needs to

evolve from today‟s heavier (six to ten pound) laptops to lighter weight, high

performance mobile PCs. Crusoe will help propel that change.‟”17

     See footnote #11
     See footnote #14
                                    Works Cited Dorsey, John. Transmeta

      Blends Hard and Software on New Crusoe Chip. (January 19, 2000). Code

      Morphing™ software. Crusoe™ Processor

      Model TM5800. LongRun® Power

      Management Technology. Performance

      Under Server Computing Workloads. Transmeta Breaks

      the Silence, Unveils Smart Processor to Revolutionize Mobile Internet Computing. Transmeta’s Crusoe: A Low-Power

      x86-Compatible Microprocessor built with Software. Transmeta Crusoe – Leaving

      Hardware. (January 2000). Transmeta Crusoe TM5x00.

To top