Crusoe Finally Reaches Civilization. Is It Too Late? Does Civilization Care?

Readers of CTR should have been rather uninspired by supersecret Transmeta Corp.'s formal introduction of its supersecret new microprocessor, Crusoe, in late January. After all, we revealed key details of this CPU in our April 1999 issue (available online at www.wwpi.com/ctreditorial.html). While at that time, the company would neither confirm nor deny our report, it turns out to have been accurate.

Crusoe is currently a family of two microprocessors--the TM3 120 and TM5400--aimed at mobile computing devices. As we reported last year, Crusoe uses so-called "morphing" software to translate chip instructions, and it features a software-based power consumption control architecture. One version of the chip will run a new, mobile-device version of Linux, while the other will support Windows. Transmeta's breathless rollout of Crusoe's variable megahertz feature was somewhat eclipsed by Intel's introduction of a similar feature in its chips a few days before (SpeedStep), but the chip still has some very interesting aspects that deserve discussion.

Morphology

Crusoe's main technological feature is Code Morphing software (hereafter CM). CM is basically microcode that sits between the Crusoe CPU core and the surrounding instructions that are fed to the processor, including application, Os, and BIOS layers. A Crusoe processor uses a Very-Long Instruction Word (VLIW) instruction set at its core; CM software is responsible for the translation of x86 instructions to VLIW so that they can be processed by the CPU.

To understand how this differs from conventional x86 processors, recall that Intel Pentium processors (as well as other, compatible CPUs) use a superscalar, pipelined architecture. In this model, the CPU hardware is responsible for fetching instructions from memory, decoding them into micro-operations, and then reordering them so they can be executed in parallel by the processor. Transmeta's CM microcode takes care of the translation and reordering process, enabling explicitly parallel instructions to be fed to a simplified processor. Thus, fewer transistors are needed in the Crusoe core for reordering processing instructions.

A second aspect of CM software is that it is able to translate groups of x86 instructions at the same time, whereas a traditional x86 processor translates instructions one at a time. These CM translation strings are kept in a translation cache, to which the system can turn when the same code is executed at another time. Transmeta says this significantly improves performance.

"Adding more functional units to a chip does not necessarily give the bang-for-buck in terms of price versus performance," says Transmeta's Doug Laird, VP of product development. "While about half of a PIII's transistors are in logic, Crusoe's are in the translation cache."

Both the translation cache and the CM software reside in separate Flash ROM that is not accessible to x86 code. For better performance, the CM software is copied from ROM to DRAM at initialization time. Interestingly, Transmeta notes that the processor's translation caching feature may result in poor benchmarks scores for Crusoe. This is because many benchmarking programs initiate a huge array of non-repetitive instructions, which Transmeta says--probably correctly--is not indicative of the way most people use applications. The company has created new benchmark tests which it believes better reflect the way consumers actually use their programs.

In The Long Run

The second aspect of Crusoe that we should examine is LongRun, a power management feature similar to Intel's SpeedStep. LongRun is a software module that can both reduce the frequency (clock speed) of the processor and its voltage, on-the-fly. LongRun is implemented in the TM5400 flavor of Crusoe, which runs at 700MHz and is targeted at lightweight portables.

Both CM software and LongRun allow Crusoe processors to run significantly cooler than comparable Pentium II and III CPUs. The TM5400 does not even have an active cooling mechanism, and the chip will run on as little as 10-20 milliwats of power, and 1-2 watts during peak usage. By contrast, mobile Pentium III chips consume about 15 watts of power during peak usage. (Note that with SpeedStep installed, this wattage drops to about 8 watts under battery power.)

And there are other differences. While both the mobile PIII and the TM54000 are built using .18 micron process technology, the former is 106mm2, while the latter is just 73mm2. Note, however, that Transmeta's processor has 128KB of L1 cache, while Intel's chip has just 32KB. Transmeta claims that LongRun and CM software allows its chips to run virtually all day on a single battery charge.

Transmeta's software-powered chips offer several compelling features, even putting aside their cooler operation and longer battery life. First, they can be upgraded through software--and hence through the Internet--as more sophisticated translation algorithms are developed and as the company learns more about the way people are actually using the chips. Second, they will--in theory at least--be cheaper to fabricate, because they use fewer transistors and are significantly less complex architecturally than competing offerings. And third, they offer a promising design for low-power, small-foot print devices which today cannot support the overhead associated with Intel CPUs.

All that said, Crusoe is by no means a perfect solution, for a number of reasons, some obvious and some not so obvious. First, the obvious. Adding increased software complexity to a CPU is not necessarily a step forward. Other companies--including DEC and Insignia--have tried similar, software-based emulation schemes and performance has shown significant degradation versus native code processing. Other, similar software-based processing solutions (like Java's JIT compiler) have proven to be hampered by poor performance when using bytecode. Transmeta's Laird points out that the Java VM and other such compilers are burdened with the overhead associated with running on top of the OS, a drawback which Crusoe eliminates.

In addition, a reliance on soft-translation opens up a world of finger pointing, as developers will inevitably find that their applications are in some way less optimized for Crusoe than for standard x86-based CPUs. Other problems may crop up at the OS level, if users find software incompatibilities with Windows.

One can also make a good case that battery drain has less to do with CPU voltage and megahertz and more to do with 15-inch LCD panels, huge hard drives, gobs of memory, and poorly designed applications. The CPU is only one small piece of a large power-conservation puzzle.

And there is the fabrication and adoption issue. Because Transmeta does not have a chip plant, Crusoe is being fabbed by IBM Microelectronics. This means that Transmeta cannot benefit from the cost economies associated with in-house manufacturing. Observant readers will remember that Cyrix suffered horribly from regular chip delays and manufacturing problems.

In addition, we live in a world where users are, unfortunately, conditioned by Intel's marketing machine. End users clamor for the fastest new CPU available, whether they need it or not. Transmeta may want to believe in a future governed by processors infused with cheap--or free--software upgrades, but this is a sales model that remains unproven. Megahertz matters; just ask Intel.

Observers, while impressed with Crusoe technology, note that Transmeta has an uphill battle in the hearts and minds of consumers. "There is absolutely room for a new architecture in mobile computing," says Gartner Group. processor analyst Martin Reynolds. "The current approach is to stuff a desktop processor into a mobile package, leading to power-limited computing." Reynolds notes, however, that Transmeta's challenge is to convince the market that these small, light notebooks are a viable option for mainstream mobile computing, rather than the sub-10 percent segment that they represent today.

Transmeta officials are convinced that their software-based processor model will result in easier and faster upgrades, equal or better performance, and longer battery life than the competition. "This is a processor that learns how you work and adapts and optimizes itself," says Transmeta's Laird. PCs based on Crusoe should be available in the second half of the year.