The Inmos Legacy
by Dick Selwood
August 2007

 

30 years ago this month, two men met at an IFIP meeting in Canada. The consequences of that meeting made an impact on the face of the electronics industry. Dick Selwood looks at Inmos and its legacy

The way Iann Barron tells it, his first meeting with Dick Petritz should have been an omen. Iann had been tasked with organising a session on the future of computers at IFIP’s conference in Toronto in 1977 and he asked Dr Petritz, who had been at Texas Instruments and was now a venture capitalist, to present a paper on semiconductors. The programme was printed, with Dr Petritz’ name in it, but Iann never heard from him until he arrived at the conference and gave his paper. During the after-session drinks with the speakers, Petritz asked Iann if he wanted to start a semiconductor company, with his role being to develop a microprocessor.

The company was eventually funded by the British Government through the (now extinct) National Enterprise Board in 1978. A site in the US covered memory design, process development and initial manufacturing, while a UK design centre in Bristol worked on the microprocessor and the UK manufacturing facility was in a Richard Rogers designed factory in Newport, South Wales.

There were always tensions. The Thatcher Government, elected in May 1979, was very unhappy at inheriting the company. Their representatives on the company’s board, the NEB-appointed external directors, knew nothing about the semiconductor industry. And there was a cultural split between the US engineers and the UK scientists, which personality clashes between Iann Barron and the third founder, memory guru Paul Schroeder, did nothing to alleviate.

The US management never understood that the UK Government wasn’t a normal venture capitalist but was expecting much more than a return on capital, looking for technology transfer to help move Britain forward in electronics technology and for thousands of jobs.

The UK press were expecting Inmos to produce a 64kbit dynamic RAM, then regarded as leading edge, and couldn’t understand that a 16k SRAM was at least as complex and potentially more profitable.

And the transputer was a mystery to everyone.

The transputer
Iann Barron’s vision was of an element that would both carry out what we would today consider a form of digital signal processing and would also, through simple interfaces, link with other transputers to produce a parallel processing array whose throughput scaled almost linearly with the number of elements. Many parallel processing architectures have extensive communication overheads, so adding a new element doesn’t always create a significant increase in power, and in some early experiments could even reduce processing available for the application.

To design the transputer, the Bristol team first created a design system. (The US memory design teams still used draughtsmen, as their effort was in tuning the memory cell and then creating an array.) With the system, a team under David May created the architecture of the transputer and Inmos also developed a programming language, occam. Both had inherent parallelism based on the theory of Communicating Sequential Processes (CSP) developed by Tony Hoare (now Professor Sir C. A. R. Hoare, Senior Researcher with Microsoft in Cambridge). To add to the complexity of the task, the team also had to design a development environment including a work station. The same team was also responsible for a graphics processing engine family.

While this was moving forward at Bristol, the US designed and launched the 16k SRAM. After a slow start it became a commercial success, with over a quarter of the market and some very good margins, helped by Intel, then the memory market technology leader, failing to ship their own device. The 64k DRAM was moderately successful in a much more crowded field, as by this time the Japanese were gearing up their fabs with DRAM and getting far better yields than many US fabs.

On the political level, the American management tried to organise a buy-out of the American operation and failed. The NEB, after nearly paying a US electronics company to take Inmos off their hands was preparing the company for a US stock market flotation or for the sale of a significant stake to an industrial consortium. Then Thorn EMI, in 1984, made an offer that not only solved the Government’s problem of what to do with Inmos, but gave them a 30 million profit on their 65 million investment and took 35 million of loan guarantees (secured on highly in-demand chip-making equipment) from the public books.

Downturn
Thorn EMI and its first appointee as Inmos chairman also knew little about semiconductors, and didn’t realise that much of the product development, including test chips for the transputer, had been held up when manufacturing capacity was used to ship as much product as possible into a boom market, making the company as attractive as possible for a floatation. When the inevitable market downturn followed, Inmos was left with few new products and it wasn’t until a year later that the first transputer chips, in the T400 family, were ready to be shipped.

Transputers were successful for a number of applications: the IBM PC used the graphics engine and many laser printers were powered by transputers. But Thorn EMI’s own problems meant that was insufficient capital to develop the transputer family properly. The T800 moved the product forward with a 64-bit floating point unit, but it wasn’t until after Thorn EMI transferred Inmos into ST’s (then SGSThompson) ownership, that the next generation, the T9000, appeared in 1992. By then it was not sufficiently powerful to provide an incentive for developers to take on what they perceived as the difficult task of learning how to programme parallel systems.

ST pulled the plug on further transputer development and Inmos staff were deployed in a variety of other activities, including more graphics controller chips. Both the manufacturing plants were sold, with Colorado Springs going first to Cray Computer and then to a succession of other companies. The Newport facility was first run as a pure play foundry with several owners and is now owned by International Rectifier.

Today Inmos doesn’t exist as an entity, but there is a significant legacy.

Legacy
The Inmos legacy has several threads. There are companies that were founded by Inmos people, there are the companies where Inmos staff have entered senior positions and there is the more intangible area of the influence that Inmos has had on the electronics industry as a whole. Also there is space, where Inmos products are still in service and Inmos ideas are exercising an influence on the development of space technology.

Spin-offs
The first Inmos spin-off was transputer based, but was a memory company. Rahul Sud, who had been active in the 16k SRAM design and then in a failed design for a 64k EEPROM founded Lattice to work on EEPROM, attracting a number of Inmos process technologists to the new company.

The first transputer company was Meiko. In the internal politics that followed Thorn’s attempts to understand and manage Inmos, a small group from the transputer team left to create supercomputers with multiple transputers. It supplied systems to major power users. In an attempt to break into the American market, it registered as an American company, and then found itself ruled out of a project for the Meteorological Office. Meiko is now part of Quadrics, itself part of Finmeccanica, still working on powerful processing equipment.

Another company created by people from the Bristol design centre was Division, exploiting virtual reality, which went public and was taken over by Parametric Technology. People from Division founded PixelFusion, now ClearSpeed, which uses massively parallel processing to create coprocessors and acceleration boards.

Motion Media (now part of Scotty Group) was another Inmos child, developing video communication networking. Their videophone technology is now available from AuPix. A spin-off from Motion Media was EsGem, who were developing a technology for tracking issues in networked products.

The Inmos family tree continues with XMOS, which has just been launched with David May, now Professor and Head of Computer Science at Bristol University, as CTO (see box).

Hotspots
All those companies are in the Bristol area, and like the other two areas where Inmos had operations, Newport in South Wales and Colorado Springs, it is now a technology hotspot. As well as the direct spin-offs, other companies have located operations in the area to fish in the pool of skilled labour, using the same dynamics as made Silicon Valley such a success.

The Bristol area has ST, obviously. Other multinationals include HP, Infineon, and Toshiba. Element 14, once part of Acorn/ARM, set up a design centre using ex-Inmos skills, which is now part of Broadcom, and some of the team from there have gone on to form Icera, a chip company working on mobile data. Independent start-ups include Elixent (now part of Matsushita), Microcosm (now part of Conexant) and, in nearby Bath, picoChip building massively parallel arrays, with wireless as a target.

South Wales has had a succession of companies setting up manufacturing plants around Newport, but has suffered from the movement to off-shore manufacturing, as to a certain extent has Colorado Springs.

In Colorado Springs, Simtek was a direct descendant, founded by Dick Petritz to concentrate on memory after he left Inmos. Microtronix, initially specialising in ferroelectrics, was another descendant, which in turn spun off Cova Technologies, Celis Semiconductor, now concentrating on RFID and Albido.

Ramtron, another ferroelectric memory company was started in Colorado Springs recruiting a significant number of Inmos people. Other companies that have been active in semiconductors in the Colorado Springs area include Atmel, Micron, Mostek, United Technologies, Honeywell and Intel, amongst others.

The roll call of major companies where Inmos alumni are in senior positions is large, as perhaps it should be, as Inmos was able to be very selective in recruitment. There are also many smaller consultancies working in the technology industries, offering such services as audio newsletters, financial consultancy, marketing, PR, or design services for semiconductors or systems. More specialist companies include Asset Recovery International, which was founded on the experience of selling the manufacturing equipment in the Colorado Springs fab, and Taeus, which helps companies with intellectual property enforcement.

Space
The transputer was welcomed in space applications, since an array of transputers could provide the very high level of redundancy that space activities need. With the long life of these projects, transputers are still in service today. For example, SOHO, the Solar & Heliospheric Observatory (sohowww.nascom.nasa.gov), a joint European Space Agency and NASA satellite, is sending back images of the sun using a transputer network.

The point-to-point technology to link transputers was formalised in the IEEE 1355 standard. This was further developed under the ESA into SpaceWire, which is now designed into a number of space missions under development in Europe, the USA and Japan. A key player in developing SpaceWire was Paul Walker from the transputer team whose company, 4Links, is now selling SpaceWire test equipment.

Another, less direct, influence can be seen in the HyperTransport standard for chip-to-chip and board-to-board communication, which is also point-to-point and was influenced, in part, by Inmos/Meiko veteran, Gerry Talbot.

Intangible
While, as we have seen, the transputer was not a lasting commercial success, it was very attractive in academic circles: a lot of doctoral theses were written in the late 1980s and early 1990s around the transputer and occam. And many of those doctoral candidates are in senior technical positions in a range of electronics companies, comfortable with the idea that parallel processing is a practicable technology. Analyst Gary Smith of Gary Smith EDA says of parallel processing: “In this area people talk almost with reverence about the transputer.”

There is still considerable research going on in the CSP field, with the WoTUG forum providing a focus for developments in CSP, including a Java implementation of CSP and an annual conference. The recent growth in multicore processors, however, doesn’t echo the philosophy behind the Inmos approach. Where Inmos developed processor and programming tools in tandem, with both chip and software using the same CSP model, the multicore approach takes existing processor architectures and tries to impose parallelism on to it. Iann Barron has said recently, “Intel has hit the wall with the quad core. All they are doing is multi-threading on several processors rather than just one; there isn’t enough inherent parallelism in Windows and current user applications to keep even four processors usefully busy.” He says that there is no magic formula to allow developers to shoehorn large applications straight into parallel processing. Instead, as applications are extended, new features should be implemented for parallelism, with those areas that lend themselves most easily to parallelism, for example graphics rendering, being reimplemented over time.

There are approaches to higher performance that are more closely aligned to the Inmos approach: a schematic of Cradle Technologies’ multiple DSP engines on a single chip looks very like an early schematic for the use of multiple transputers at a board level. And ClearSpeed has 96 processing engines on a single die to produce the power of some supercomputers when Inmos was founded. While Inmos failed in persuading most people that it was possible to create parallel processing systems that are easy to programme, the ideas are refusing to die. As the demand for more and more performance continues, it will be interesting to see the Inmos ideas re-invented with new names.

XMOS – Keeping the Faith
Software Defined Silicon (SDS) from XMOS is a reinterpretation of the transputer ideas for the 21st century.

David May, CTO of XMOS, was the architect of the Inmos transputer and is currently professor of computer science at Bristol University, where he has developed the two elements of the XMOS SDS approach, silicon and a software design flow. SDS aims to provide consumer electronics developers with silicon that has ASIC flexibility and prices ($1 per chip) and is at least as easy to programme as an FPGA.

The silicon is an array of simple, event-driven processors, the XCores. Their functionality is defined in the development system using XMOS C language (XC) which is standard C with extensions to cover, for example, IO port definition, thread behaviour and parallelism. The developer can define dynamically how the processor matches processing resources to each task, balancing control processing, signal processing and IO. Standard C/C++ is used for applications, allowing re-use of legacy code, and a soft IP library is being developed.

First silicon is in 90nm fabrication at TSMC. Later this year there will be more information on the architecture and a beta release of the tool set, which also includes compilers, a debugger and development boards. Formal release of tools and silicon is planned for the first quarter of 2008.