Lexra

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

~ the obit ~

Four instructions killed this company. The MIPS I instruction set has a group of unaligned load and store operations, written lwl, lwr, swl and swr, and those four were covered by United States patent 4,814,976. Lexra built processor cores that ran everything else in MIPS I and deliberately implemented none of those four. Hit one of them on a Lexra core and you got a reserved instruction exception, the same thing you got from any opcode that did not exist.

The company was founded in 1997 in Waltham, Massachusetts by Charlie Cheng, who was chief executive for all six years, and Pat Hays. It sold synthesizable cores rather than chips: a customer licensed the design, dropped it into their own silicon, and shipped it inside something else. The first was the LX4080, and the line grew through the LX4180 with its compressed MIPS16 encoding and on-chip debug, the dual-issue LX4280, the LX5280 with the Radiax signal-processing extensions, and the multithreaded LX8000. Jonah Probell, who wrote hardware there, counted roughly forty licensees and about thirty that actually manufactured chips.

The legal argument was never that the patents had run out. Lexra's position was that an instruction set is not a thing anyone can own, that the 976 patent reached only those four operations, and that a design which refused to execute them was therefore clean. Software had to meet it halfway: the Linux MIPS kernel took GPL patches that trapped the missing opcodes and emulated them with shifts and aligned accesses, and the other route was to teach gcc never to emit them. Probell put the share of MIPS development tools that ran on a Lexra core untouched at about ninety-five percent.

Silicon Graphics got there before MIPS did. In April 1998 SGI sued over the way Lexra advertised itself, on false advertising and trademark grounds, and that ended in October 1998 with a memorandum requiring Lexra to say in public that it did not implement the full MIPS I instruction set. The real case came on 28 October 1999, when MIPS Technologies filed in the Northern District of California on the 976 patent and a second one covering overflow handling, asking for an injunction and enhanced damages. Lexra counterclaimed on 12 November 1999, arguing the infringement claims had been brought in bad faith.

~ the rap sheet ~

~ what the court never decided ~

Lexra was winning on the patent itself. It filed three separate reexamination requests at the patent office, and around June 2001 the examiner preliminarily knocked out six of the fourteen claims in the 976 patent on prior art from a 1975 IBM filing. Lexra had only disputed four of them; the office took two more on its own. MIPS appealed that and got a broad reading of the patent language from the judge, and a claim construction ruling landed on 14 September 2001 after a Markman hearing that April.

Then it settled, in December 2001, with the announcement going out on 31 December. Probell says the settlement came one week before trial and that dropping the reexamination request was one of the conditions. Lexra conceded the patents were valid, agreed to stop licensing processor cores, and took a MIPS32 architecture license. MIPS bought the relevant technology for $6.0 million in cash plus about $400,000 in costs, on top of a $1.4 million equity stake it already held, and by November 2002 Lexra's own homepage described the company as a licensee of the architecture it had spent four years working around. Probell's summary of the law is the line to keep: there is still no precedent holding that a processor which runs MIPS I and treats unaligned loads and stores as reserved instructions infringes the 976 patent. MIPS then moved every one of its 32-bit cores to MIPS32, extending the old instruction set with freshly patented features, which closed off the clean-room route for whoever tried it next. The 976 patent itself expired on 23 December 2006.

~ the core in the router ~

The cores shipped widely enough that the argument was never academic. Zoran's Vaddis 5 put one inside DVD players that shipped under the Toshiba and Pioneer names. Macronix, Micronas, Trident Microsystems and Analog Devices all took licenses. Realtek built the RTL8181 and two further parts on the LX5280, and those went into the cheap wireless access points that hobbyists were still reflashing years later.

That last one accounts for nearly every person who ever touched this company. A plastic router from a computer shop, the kind that cost less than the service it connected you to, ran an instruction set with four holes in it and a Massachusetts lawsuit attached. Nothing on the box said so.

~ the takeover that nearly happened ~

The archived homepage from 25 November 2002 promises a first product in the fourth quarter of that year. It never came: the LX4580 was licensed by nobody, and the chip it was meant to go into never reached tape-out. What Lexra spent those months on instead was an attempt to buy MIPS. Lexra pulled together a group of unhappy MIPS licensees, a J.P. Morgan banking group, a major foundry and outside financiers, with a plan to offer something north of $4 a share against a stock trading at $1.25. MIPS climbed to $3.50 by December 2002 and the margin disappeared with it, so the group gave up. That was the same month MIPS wrote its Lexra equity down to nothing, having concluded that the chance of Lexra continuing as a viable business was remote. It also took a $1.2 million impairment on the technology it had bought, on the grounds that the technology was not compliant with its other products.

On 3 February 2003 EE Times reported Cheng confirming the company was dissolving its assets, with its thirty-five employees already told. MIPS put the ending on the record in its own annual filing that September, in one sentence: Lexra ceased operations in the quarter ended March 31, 2003. Wikipedia says January and cites nothing for it. lexra.com now runs on nameservers belonging to a brand marketplace, which lists the word Lexra for sale.

@}-->--

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

« previous  ·  random  ·  next »