Transmeta: hype and processor performance

Transmeta was founded to design and develop an Intel compatible microprocessor/software combination. Transmeta was founded by David Ditzel, who was one the main computer architects of Sun Microsystem's SPARC processor. In addition to Ditzel, Transmeta hired a number of other "big names". These included Steve Johnson who did early C compiler work at AT&T, Linus Torvalds, the original author and force behind the Linux operating system and Dave Taylor, one of the key developers of the Quake computer game. Dave Taylor ported Quake to Linux. In hiring Torvalds and Taylor it appeared to some people that Transmeta was hiring for name recognition first and skills that could be contributed to the company second. This is not to suggest that these individuals are not extremely talented. But it does beg the question of whether people with the same background and skills would have been hired if they were not well known in the industry.

When Transmeta was founded Intel was not concentrating heavily on the market for processors for portable devices. The profit margins in this market were thin, compared to what they could get in the desk top market. Transmeta was founded as a "fabless" semiconductor company that would design a processor that was faster and cheaper than the Intel processors for laptop systems. The Transmeta processor would, however, be fully compatible with the Intel processor (the so called x86 instruction set).

Intel is infamous for suing any competitor that produces a processor that executes a compatible instruction set. For example, they had a long running suit against Advanced Micro Devices. To avoid such a suit Transmeta designed their processor to run in an environment where software translated the Intel instruction set into Transmeta processor instructions. This is sometimes called "on the fly" code translation. Despite Transmeta's claims to inventing this idea, a significant body of work was done first at IBM.

There is an obvious problem with on-the-fly translation: instruction execution is slowed down by going through software translation. In contrast, instruction execution in VLSI is implemented by hardware components that attempt to make use of parallelism and pipelining. At the time Transmeta claimed that since their processor would be simpler than the mind bendingly complex Pentium processors from Intel. Since the Transmeta processor would be simpler, the company claimed, it could run at a higher clock rate (shorter critical paths in the CPU), offsetting the overhead of software translation. The Transmeta processor would also use caching so that portions of the code that had already been translated would not have to be translated when executed again.

Transmeta was not able to achieve their goal of a faster, cheaper processor for laptop computer applications. So they repositioned the processor as a low power device for portable computer applications. There were problems here as well. The most lucrative market for portable devices is laptop computers. When the power consumption of a laptop is examined, only a fraction is actually consumed by the processor. The largest power consumers are the hard disk and the display. Memory is also a significant consumer of power. For this reason, laptop manufactures use slower, less power intensive, memory, which has a negative impact on the performance that can be delivered by the processor. In terms of power consumption, the processor comes in around third or fourth place.

The fact that the processor is not an overriding consumer of power in a laptop means that the Transmeta processor is not likely to tempt any laptop manufacturer to switch from Intel. The only applications left are tablet or PDA devices, where the processor does consume a significant fraction of the power. Transmeta has had some design wins for PDA devices sold in Japan, but at the time of this writing, there have been few design wins for devices sold in the United States.

The market for low power processors for portable computing devices is hotly competitive. For over five years chip design tool companies like Synopsys have worked to develop software tools to help design lower power chips. Intel has gone from ignoring this market to shipping low power versions of the Pentium. Both Intel and AMD have become very price competitive and the margins have fallen for low power processors. All of this has removed any real advantage that Transmeta had. What is left is the hype. But hype will not deliver quarterly profits.

A Wired News article reports that Transmeta is being sued by the usual gang (e.g., Milberg Weiss) in response to their stock tanking. What is interesting is that this suit claims that Transmeta was not fully forth coming about the issues surrounding their processor in their IPO.

In the best of times the high tech environment is risky and I tend to think that Milberg Weiss suits are usually, as they say, "without merit". In many cases Milberg Weiss has acted as a blackmailer, threatening long term costly litigation to force companies to settle, even when Wilberg Weiss' claims have been weak. However, in this case Milberg's argument does not seem to be based on simple blackmail. In the filing for the suit, Milberg writes:

The Prospectus was false and misleading because it portrayed the Company's Crusoe technology as being able to "simultaneously offer long battery life, high performance and x86 compatibility." In fact, the Company was not providing revolutionary technology, but rather, Transmeta's technology was significantly slower than comparable Intel and AMD products. Despite this fact that the Crusoe technology was anything but "high performance," the Prospectus falsely stated that:

We develop and sell software-based microprocessors and develop additional hardware and software technologies that enable computer manufacturers to build computers that simultaneously offer long battery life, high performance and x86 compatibility

Transmeta ran extensive internal benchmark tests so it seems likely that they knew about the performance of the Transmeta processor. They did disclose that the processor would run some instructions slower, but they still claimed that overall performance would at least be comparable to Intel's. It would be interesting to know if their internal benchmarks bear this out.

The Milberg Weiss web page on this suit has moved from my original link or been removed entirely. A summary of the filing can be found on a Stanford Law School web page. I have not found any information on the current status of this suit. Perhaps Milberg Weiss has dropped it. In bringing this suit, Milberg Weiss would be under pressure to collect some money before Transmeta goes bankrupt. Since Transmeta's clock seems to be ticking faster than that of the court process, Milberg Weiss may simply have concluded that they would not make a profit.

Asynchronous Processors

Although Transmeta is built largely on hype, this does not mean that there are not low power approaches to higher performance processors. One of the most promising is asynchronous design. An asynchronous processor would not have a clock. Rather, the logic would be self timed. An overview of this technology can be found in the Technology Review article It's Time for Clockless Chips

The problem with asynchronous design is that all of the design tools that currently exist and all of the simulation tools rely on synchronous (clocked) design. So a great deal of work remains to be done to make this approach practical.

The Transmeta Death Watch

The November 5, 2001 article referenced below seems to have triggered a Transmeta death watch. The speculation is that Transmeta will either go under or get bought by a company like VIA Technology. At the time of this writing, Transmeta does have quite a bit of money in the bank, so they may still have space to pull a rabbit (or more hype) out of their hat.

As this page is updated in July 2002, Transmeta's future does not look any brighter than it did in November of 2001. Transmeta has consumed big chunks of their cash reserve. The company announced that they are firing about 40 percent of their staff. As the value of the Transmeta hype recedes, the fact that they do not have technology that delivers any important advantage is becoming obvious to more people. Intel has not ceded the low power market and has been aggressively marketing lower power processors. Low power processors also exist based on the ARM and MIPS instruction set architecture. Other technologies, like asynchronous design, promise low power high performance devices in the future.

In November of 2002, as this article is updated, Transmeta still survives, but continues to lose money. Transmeta's stock has fallen from its IPO open of $21 about 95% to $1.15. The company currently has $150 million in cash and is burning about $20 million per quarter. According to Transmeta's current CEO, the company hopes to gain profitability (presumably on a cash flow basis) in the later part of 2003. That is, if they are not eaten by Intel first. AMD has been struggling against Intel with a much better product line.

The semiconductor industry is currently going through its worst down turn in history (many are referring to it as a depression). In such an environment a buyout is less likely to save Transmeta (assuming that there is actually something there worth buying).

Web References

These references are ordered from newest to oldest.

Ian Kaplan, June 2001
Revised: January 2005