Transmeta

» see transmeta.com as it lived, on the wayback machine

~ the obit ~

For most of its life Transmeta was a rumor with a payroll. A website went up around 1997 carrying one line, "Not yet the Transmeta home page", and it stayed close to that for two and a half years while the company built something nobody outside the building had seen. Linus Torvalds worked there. Paul Allen had money in it. The silence did more marketing than a launch campaign could have managed, and by the time the doors opened the guesses had gotten silly enough that one reporter joked in print about alien technology.

The reveal came on 19 January 2000 at an event in Saratoga, California, and the product was called Crusoe. Seven people had founded the company in Santa Clara in 1995, with Dave Ditzel out of Sun Microsystems running it. What they put on the table was a processor with very little x86 hardware on the die. The x86 part lived in a software layer they called Code Morphing Software, which read Intel instructions and turned them into the long native words the silicon underneath actually executed.

The argument was power. A chip that does less in hardware burns less, and Transmeta put battery numbers next to its part that no Intel mobile processor could answer in 2000. Sony built the VAIO PictureBook around it and the tiny subnotebooks of that period used it because nothing else fit inside the thermal budget. The IPO landed on 7 November 2000 at 21 dollars a share and touched 50.26 the same day, the last large semiconductor listing of the dot-com bubble.

The machines Crusoe went into were subnotebooks, and a subnotebook in Turkey in 2001 was not something you bought at a price that made sense. A home computer here was a parts list read out across a shop counter and carried home as a tower. My machine sat on a desk and stayed there. How many hours it could run away from a wall socket was not a number anyone in that room had a reason to ask about, which means the one thing Transmeta was better at than Intel was the one thing I had no use for.

~ the rap sheet ~

~ what code morphing actually did ~

The hardware was a VLIW core, 128 bits wide on Crusoe and 256 on the later Efficeon, with an instruction set that no compiler on earth targeted. The translation layer loaded out of flash before the operating system did, took x86 instructions apart, and packed them into native bundles. Translations got cached, so a loop that ran ten thousand times was converted once and then reused. It also meant the processor could be corrected after it shipped, because a bad translation was a software bug and a software bug can be patched.

The cost showed up on every instruction the chip had not seen yet. Code that sat in a tight loop ran well. Code that jumped around, which describes most desktop software, spent its time being converted instead of run. The distance between those two cases is where every review of the thing ended up.

~ the benchmark problem ~

The analyst Tom Halfhill put it at roughly a 500 MHz Pentium III worth of work out of a 700 MHz Crusoe. Transmeta's answer was that the comparison left out the power budget, which was fair and did not help, because the number printed on the card next to the laptop in the shop was megahertz.

Efficeon, announced on 14 October 2003, was the correction. Twice the work per clock, most of the gap closed. Intel had already shipped the Pentium M that March. Centrino turned battery life into a brand name and did it with a processor that ran x86 in hardware and won the benchmarks as well. After that there was nothing left for Transmeta to be better at.

~ the settlement that paid the bills ~

In January 2005 the company announced it was no longer a semiconductor product company. Most of the staff went, the fabless chip business was wound down, and what remained licensed the patents to whoever wanted them. This is the phase where a hardware company becomes a filing cabinet.

On 24 October 2007 Intel settled the patent suit between them for 150 million dollars up front plus 20 million a year for five years. AMD put 7.5 million into the company the same year. So the outfit that had burned close to 600 million dollars trying to get Crusoe into the world closed out by collecting a quarter of a billion from the company that had beaten it. That money is what Novafora was buying.

~ who owns the patents ~

Novafora, a video processing startup, announced the acquisition on 17 November 2008 for 255.6 million dollars in cash and closed it on 28 January 2009. A week later, on 4 February, the patent portfolio was sold on again to Intellectual Ventures. Novafora collapsed that July. Transmeta had 24 employees at the end.

What is left is a license pool that anyone can pay to use. The idea did not die so much as get filed. Dynamic binary translation is how every x86-on-ARM compatibility layer shipping today works, and the emulation layer that lets an Apple laptop run Intel software is doing the same trick with better hardware underneath it. Transmeta paid 600 million dollars to survey that ground and then handed the map to a licensing company.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »