Transmeta

Transmeta Corporation

Company type	Public
Industry	Semiconductor
Founded	1995
Founder	Bob Cmelik, Dave Ditzel, Colin Hunter, Ed Kelly, Doug Laird, Malcolm Wing, Greg Zyner
Defunct	2009
Fate	Acquired by Novafora; patent assets sold to Intellectual Ventures
Headquarters	Santa Clara, California , United States
Area served	Worldwide
Key people	David Ditzel (CEO 1995–2001), Mark Allen (CEO 2001), Murray Goldman (CEO 2001–2002), Matt R. Perry (CEO 2002–2005), Art Swift (CEO 2005–2007), Lester Crudele (CEO 2007–2009)
Products	Crusoe processor, Efficeon processor, Code Morphing Software, LongRun and LongRun2 power management technologies
Revenue	$2.48 million (2007)
Number of employees	24 (2009)
Parent	Novafora (from 2009)

Transmeta Corporation was a fabless semiconductor firm based in Santa Clara, California, United States. Established in 1995 by founders Bob Cmelik, Dave Ditzel, Colin Hunter, Ed Kelly, Doug Laird, Malcolm Wing, and Greg Zyner, the company specialized in creating low-power microprocessors compatible with the x86 architecture. These chips utilized a VLIW core architecture combined with a proprietary software layer known as Code Morphing Software (CMS).

The CMS technology integrated an interpreter, runtime environment, and dynamic binary translation system to enhance execution of x86 instructions by analyzing usage patterns and incrementally producing optimized translations. The underlying VLIW hardware incorporated features such as speculation support, memory alias detection, and accommodation for self-modifying code, all contributing to improved CMS efficiency and reduced power consumption. This innovative design catered to sectors requiring minimal energy use and heat output, including subnotebooks, laptops, desktop PCs, blade servers, tablet devices, and silent computing systems.

Transmeta introduced its initial product, the Crusoe processor (internally called 'Fred'), in January 2000. Later that year, the company held its initial public offering in November. The subsequent major release was the Efficeon processor ('Astro'), which arrived in October 2003. By 2005, Transmeta pivoted towards monetizing its intellectual property by licensing microprocessor patents and semiconductor innovations to bolster financial results. Following workforce reductions in 2007, it fully transitioned away from manufacturing operations to focus exclusively on IP licensing.

In January 2009, Novafora acquired Transmeta; simultaneously, its patent assets were sold to Intellectual Ventures, which manages non-exclusive licensing of the technology. Before this acquisition, Transmeta had licensed critical technologies—including the LongRun and LongRun2 power management solutions—to prominent industry players such as Intel (under a perpetual non-exclusive agreement), Nvidia, Sony, Fujitsu, and NEC. These collaborations extended Transmeta’s market reach and revenue opportunities despite the challenges faced in semiconductor fabrication.

At its operational peak around 2009, Transmeta employed approximately 24 personnel and reported revenues of $2.48 million in 2007, while incurring operating losses exceeding $61 million. The company’s distinctive edge stemmed from its combination of VLIW hardware with adaptive software that enabled energy-efficient execution of x86 code—an approach unique within the semiconductor sector that influenced future designs for low-power processors.

Stealth start-up

Founded in 1995, Transmeta functioned as a stealth startup until publicly unveiling itself on January 19, 2000. Throughout this confidential phase, it secured more than 2,000 non-disclosure agreements to maintain strict secrecy over its product development and corporate strategy. From mid-1997 until early 2000, the company’s website remained sparse, displaying only an enigmatic statement alluding to its cutting-edge mobile processor technology.

The primary focus during this period was the development of the Crusoe processor—engineered specifically for mobile computing—with an innovative architecture that translated x86 instructions into native very long instruction word (VLIW) commands. This approach positioned Transmeta as a competitor within the semiconductor market by offering processors that promised lower costs alongside high performance tailored for mobile devices.

Although public details were scarce during these formative years, Transmeta’s stealth mode effectively safeguarded its intellectual property while generating anticipation for its forthcoming products and business model.

Public introduction

On January 19, 2000, Transmeta unveiled its x86-compatible processor, named Crusoe, at an event held in Saratoga, California. Alongside the announcement, the company published an 18-page whitepaper explaining the underlying technology.

Transmeta marketed its microprocessor as a breakthrough in the low-power segment, targeting leadership in both energy efficiency and computational performance. Nevertheless, early assessments revealed that Crusoe's actual performance fell short of expectations. Concurrently, competitors Intel and AMD enhanced their processors’ speed and power efficiency, restricting Crusoe's appeal to a niche market focused on low-volume, small form factor (SFF), and low-power applications.

The company went public on November 7, 2000, with an initial public offering priced at $21 per share. Shares surged to $50.26 before closing at $46 on the first trading day—positioning Transmeta among the final significant tech IPOs during the dot-com bubble and outperforming until Google's public offering in 2004.

Encountering operational difficulties, Transmeta initiated workforce reductions in July 2002, cutting approximately 40% of its employees. In October 2003, it launched the Efficeon processor, which was claimed to deliver twice the performance of Crusoe at matching clock speeds. However, increased complexity and higher power consumption diminished its competitive edge.

By January 2005, Transmeta pivoted from producing semiconductor products to focusing exclusively on licensing intellectual property. In March of that year, it laid off 68 staff members but maintained a core team of 208 employees; Sony emerged as a principal licensee utilizing Transmeta’s LongRun2 power management technology.

On May 31, 2005, Transmeta entered into asset purchase and licensing agreements with Culture.com Technology Limited based in Hong Kong; yet, delays in securing export licenses led to the agreement’s termination in February 2006.

The company reported its first profitable quarter on August 10, 2005. In March 2006, it was disclosed that Transmeta was collaborating on a Microsoft initiative—a secure computing platform branded by AMD intended for Microsoft’s FlexGo program.

On October 11, 2006, Transmeta initiated legal action against Intel for infringement of ten patents pertaining to computer architecture and energy efficiency technologies embedded in various Intel product lines.

On February 7, 2007, Transmeta ceased hardware development by shutting down its engineering services division to concentrate solely on intellectual property licensing. Following this shift, AMD invested $7.5 million to access Transmeta’s patents related to energy-efficient technologies.

On October 24, 2007, Transmeta reached a settlement with Intel whereby Intel paid an upfront sum of $150 million plus annual payments of $20 million over five years and obtained licenses for several patents from Transmeta. As part of this agreement, Transmeta consented to discontinue manufacturing x86 processors. The deal also included roughly $34 million paid to three executives—a matter that drew criticism from investors. Intel later consolidated these annual payments into a single lump sum during 2008.

On August 8, 2008, Transmeta licensed its LongRun and other low-power technologies to Nvidia for a one-time payment of $25 million. Subsequently, on November 17, it agreed to be acquired by Novafora for $255.6 million in cash; this transaction was finalized on January 28, 2009.

Transmeta’s patent portfolio was acquired by Intellectual Venture Funding LLC on February 4, 2009.

By late July 2009, Novafora collapsed due to financial challenges and ineffective execution.

Corporate governance

Throughout its existence, Transmeta was managed by six different chief executives:

CEO	Tenure
David Ditzel	1995–2001
Mark Allen	2001
Murray Goldman with Hugh Barnes as COO	2001–2002
Matt R. Perry	2002–2005
Art Swift	2005–2007
Lester Crudele	2007–2009

Notable employees

Among Transmeta’s workforce were distinguished professionals such as Linux originator Linus Torvalds, kernel developer Hans Peter Anvin, Yacc creator Stephen C. Johnson, and game programmer Dave D. Taylor. Their involvement enhanced the company’s visibility within the industry and bolstered its public relations efforts.

Financial history

Between 1996 and 2007, Transmeta experienced variable financial results including changes in revenue, operational costs, gross margins, and net losses. Although Upside magazine acknowledged it as a key Silicon Valley enterprise, the company did not attain profitability as a semiconductor manufacturer.

Funding

Over its operational period, Transmeta raised close to $969 million in capital.

Year	Quarter	Amount ($ million)	Description
1996	–	288	Initial capital injection
2000	Q2	88	Additional venture funding
2000	Q4	273	Proceeds from initial public offering
2003	Q4	83	Secondary stock offering
2007	Q2	7.5	Investment by AMD
2007	Q4	150	Settlement payment from Intel
2008	Q3	80	Additional settlement funds from Intel

Crusoe

Transmeta’s first microprocessor line, named Crusoe after Robinson Crusoe, initially generated interest but was criticized for modest real-world performance improvements and only slight power efficiency advantages over competing Intel and AMD chips. Its proprietary Code Morphing Software (CMS) enhanced benchmark scores but introduced processing overhead that reduced efficiency in demanding applications. Furthermore, the processor’s simplistic VLIW architecture and limited southbridge bandwidth restricted graphics and input/output capabilities, leading to concerns about its complete x86 compatibility.

Efficeon

The Efficeon processor represented Transmeta’s second-generation 256-bit VLIW design, prioritizing computational efficiency alongside low power consumption and thermal control. A version running at 1.6 GHz on a 90 nm fabrication process matched the performance and power profile of Intel’s contemporaneous 1.6 GHz Atom (45 nm), while benefiting from an integrated Northbridge that lowered energy usage.

Efficeon overcame many of the Crusoe’s shortcomings by delivering approximately twice the practical performance and featuring a smaller die size—68 mm² compared to the Pentium 4’s 112 mm² at equivalent process technology. Nonetheless, reviews frequently overlooked its intended thermal limits and application focus, resulting in unfavorable comparisons with higher-power processors such as Intel’s Pentium M (Banias) and AMD’s Mobile Athlon XP, where Efficeon’s performance was comparatively weaker.

Technology

Transmeta processors integrated in-order very long instruction word (VLIW) cores with a dynamic binary translation software layer to deliver compatibility with the x86 instruction set. The company branded this innovation as "Code Morphing," naming the software component Code Morphing Software (CMS).

To achieve significant power savings—approximately 2.5 times lower consumption—Transmeta employed reverse body bias, a power reduction technique also utilized in XScale processors.

Code Morphing Software

Code Morphing Software (CMS) allowed Transmeta’s microprocessors to run x86 instructions by converting them into their proprietary VLIW instruction format. CMS comprised an interpreter, a runtime environment, and a dynamic binary translator that analyzed execution patterns and heuristics to optimize code translation quality, surpassing similar technologies like Shade.

Although earlier comparable technologies existed during the 1990s—such as Wabi (software) for Solaris/Linux, FX!32 for Alpha, and the Mac 68K emulator—Transmeta’s solution provided comprehensive x86 instruction compatibility from system boot through advanced multimedia instructions. This comprehensive support offered a strategic edge in market adaptability and product versatility.

CMS functioned similarly to compiler optimization passes: it initially translated x86 instructions straightforwardly, then refined and combined them into efficient VLIW "molecules" capable of executing faster than native x86 code.

Transmeta emphasized several commercial advantages of CMS:

1. Accelerated product updates delivered via software without requiring hardware redesigns, enabling swift adaptation to Intel and AMD instruction set enhancements.
2. Performance and power consumption optimizations tailored through software adjustments according to market demands.
3. Software-based remedies addressing hardware design or manufacturing defects.
4. Ability to concentrate on core improvements and power efficiency without being limited by legacy x86 architecture constraints.
5. Capability to emulate multiple architectures concurrently, exemplified by running pico-Java alongside x86 on the same hardware platform.

Despite early speculation about hybrid PowerPC/x86 processors, Transmeta initially concentrated on the low-power x86 segment.

The practical benefits of CMS were demonstrated through software upgrades deployed in the field: for example, in 2002, CPU performance was improved on the HP Compaq TC1000 tablet PC without any hardware modifications; similarly, in 2004, support for NX bit and SSE3 was incorporated into the Transmeta Efficeon series via software updates. Nonetheless, such upgrades were infrequent due to vendor concerns regarding post-sale support costs and quality control.

VLIW core

Combined with its code morphing software, the Efficeon processor closely resembled features found in Intel Pentium 4 CPUs while incorporating a fully integrated memory controller, a HyperTransport I/O bus, and support for the NX bit, an x86 no-execute extension introduced with CMS version 6.0.4. Although its computational throughput was generally lower than mobile processors such as the Intel Pentium M, detailed performance comparisons remain limited.

Efficeon chips were available in 783- and 592-contact ball grid array packages with relatively low power consumption—down to 3 watts at 1 GHz and up to 7 watts at 1.5 GHz—allowing for passive cooling solutions. Two generations were produced: the initial TM8600 utilized a TSMC 130 nm fabrication process with clock speeds reaching 1.1 GHz; the subsequent TM8800 and TM8820 employed a Fujitsu 90 nm process with frequencies ranging from 1 GHz to 1.7 GHz.

Internally, the Efficeon architecture featured dual arithmetic logic units, load/store/add units, execution units, floating-point/MMX/SSE/SSE2 units, as well as branch prediction, aliasing, and control units. Its VLIW core executed a single 256-bit instruction per cycle—termed a molecule—comprising eight individual 32-bit instructions or atoms.

The cache hierarchy included an on-die 128-KB L1 instruction cache, a 64-KB L1 data cache, and a 1-MB L2 cache. Additionally, the code morphing software allocated approximately 32 MB of main memory for caching dynamically translated x86 instructions.

Native compilation

Although it was theoretically feasible to optimize x86 code for Code Morphing Software or to directly target the native VLIW architecture, experts in the field, including Linus Torvalds, deemed this approach impractical due to inherent architectural constraints such as the absence of memory protection and a missing MMU for code access management. Transmeta refrained from providing detailed information necessary for native compilation, citing concerns related to the security risks of exposing microcode interfaces.

Further reverse engineering efforts in 2004 uncovered fundamental limitations that hindered operating systems like Linux from being adapted to the native VLIW architecture. Additionally, comparative studies raised doubts about certain patented technologies claimed by Transmeta, highlighting potential conflicts with earlier patents held by IBM and suggesting possible issues regarding their patent validity.

External links

• Archived company page