Published in the November 2003 issue of IEE Review, this feature looks at the reasons why chip designers will grapple with multi-gigabyte ASCII text files and have only just begun to move to more efficient database formats.
For an industry working at the cutting edge of technology, it seems strange that the tools silicon chip designers use today handle huge volumes of data in a way that is not just clunky and inefficient but has not been seen in the mainstream computing world since the 1970s.
Where the bulk of the business computing world has been using database management systems for more than 30 years, and they began to cross over into the mechanical-CAD world in the late 1980s, the electronic design automation (EDA) world has struggled on with data formats that are little more than enormous text files to hold all the information needed to create a chip. EDA vendors tend to describe the files their tools use as databases but those flat-file system lack the sophisticated access techniques that the programmers of even simple accounting systems take for granted.
To deal with each file, an EDA tool has to read all the way through the file, build its own representation of the data internally and then work with that. When it is finished, it writes out another big text file. With multimillion-transistor chip designs now the norm, these files have grown astronomically.
According to Albert Stritter, vice-president of design automation at Infineon Technologies, by the time a 60 million transistor chip design has acquired all the information needed to have it manufactured in the fab, the file needed to describe it takes up 3Gbyte of disk space. And many tools will have waded through that data, adding in the pieces of information that will ultimately render a manufacturable chip.
"The size of the files has reached a physical limit," said Stritter. So, now the chipmaking companies are preparing themselves for the widespread use of database-driven tools.
A central database would do little to reduce the amount of data needed to describe a chip but it would dispense with the need to read all the data in and write it all back out again whenever a change needs to be made. The database management system takes care of all the accesses, fetching only those bits of data that are relevant.
Why has the industry persisted so long with such a primitive approach to managing multi-million transistor and multi-million dollar projects? The answer lies in the delicate balance of power that exists between the big EDA tools vendors and their customers. The EDA vendors have tried to introduce database-driven tools in the past and failed because customers had no desire to be locked into one vendor's particular flavour of database. Text files are inefficient to use but they are easy to reverse engineer: a useful feature whenever a vendor introduced a text file format that only its tools were designed to work with.
The situation with databases only began to change when Avant decided it would base all its tools around a proprietary database the company called MilkyWay. Like all the database-driven systems before it, it was a closed system. However, Avant was able to sell tools based around MilkyWay because, at any stage, the tools were able to read and write conventional text files that could be handled by other vendors' tools.
The two largest EDA companies have based their latest generations of tools around database management systems. Cadence Design Systems has developed one internally called OpenAccess. After buying Avant in 2002, Synopsys decided to convert most of its tools to use MilkyWay.
However, nothing has met Stritter's needs as the database that can underpin Infineon's chip-design projects. "We are looking for an open database but there is nothing on the market [that meets our needs]," he said. The problem is that the existing databases are failing to keep up with chip designers. Stritter said existing databases do not account for designs that have multiple supply voltages, yet this is an attribute that many 90nm chip designs will need to keep power consumption under control.
Some chipmaking companies are considering using a single vendor's tools throughout individual chip design projects because of the problems caused by designing at tight-geometry processes such as the upcoming 90nm technologies. Jean-Loup Lachese, EDA technical officer for Texas Instruments' wireless business unit, said: "We are now using a flow where everything is integrated because we are now working on 90nm [process] projects." However, Lachese said TI has hedged its bets on which vendor to use: "It is a risk to use just one EDA vendor. So we are working mostly with both Synopsys and Magma [Design Automation] tool-suites."
Ralph von Vignau, director of the platform infrastructure department at Philips Semiconductor, agreed: "We think opting for just one vendor is dangerous."
Although, as big customers of the EDA vendors, the chipmakers have no wish to get locked into one particular offering just because that vendor's products happen to meet their needs at one point in time, they are keen to devolve more tools development work to the EDA companies. Most of the big chipmakers have amassed portfolios of specialist tools because they perceive the EDA vendors as moving too slowly to meet their needs. This is a problem that Stritter faces on a day-to-day basis, and has led to the company developing its own techniques for handling multiple on-chip supply voltages and the complex physics of transistors built using sub-100nm processes. "We are doing what the vendors should be doing. The EDA vendors are lagging behind a little bit," he said.
The EDA vendors are meeting customers like Infineon half-way. By promising to open up their database formats, they hope to convince Lachese, Stritter and others that they will not lock customers into proprietary systems. As long as other vendors write their tools to be able to use those databases without a lengthy text-file conversion, that should counter any lock-in attempts.
By providing better information and open interfaces, the EDA vendors hope to be able to bring their customers' tools under the wing of their favoured database, even if the vendors cannot write the tools themselves.
The chipmakers are not averse to the idea. But they have one key requirement: the ability to extend and customise the database. This would let each company implement, in a custom data model, support for the details that it considers important.
Pierre Bricaud, director of research and development for the Synopsys intellectual property and systems business unit, said: "It is very difficult to standardise on the data model because each company is different." But, he said that that it would be enough to support custom tools and custom data models to have a flexible common database with an applications programming interface (API).
Cadence has gone further with its OpenAccess database. It donated the database code to the Silicon Integration Initiative (SI2) consortium for the chipmaking industry. The consortium now provides the source code to OpenAccess under a community-source licensing programme similar to the one employed by Sun Microsystems for its Java language, rather than the Gnu- or Linux-style licences, although SI2 refers to the licence model as "open source". Michael Sanie, groupe director of strategic third-party partnerships and initiatives at Cadence, said: "Open source makes OpenAccess extensible. It makes it easier to write your own tools."
Stritter said: "Open source is very important to us as [without it] we are locked into one vendor."
Lachese agreed: "It is easier to work with open-source formats." TI has adopted other de facto industry standards to replace older internal formats because the code was offered under a community- or open-source licence. One example was the Synopsys Liberty format used to describe the libraries of basic logic gates used in ASIC design. TI moved to use Liberty after Synopsys offered the code for Liberty on a community-source basis.
Although Synopsys has not provided MilkyWay code under a community- or open-source licence and currently has no plans to do so, Bricaud said: "Synopsys is very open to the open-source format." The company has also agreed to make sure its database will communicate with OpenAccess-based systems through a project called GoldenGate.
Sanie argued that the adoption of a common database by the chip designers will help foster the development of tools by small, specialist companies and academic researchers because of the support infrastructure it offers through utilities and common APIs. "There will be new tools from new vendors who download the [OpenAccess] code and write to it."
He gave the example of PDF Solutions, which started out providing consultancy on yield improvement to chip designers but has shifted to being a tools provider, using OpenAccess to provide the data to the yield-analysis software developed by the company. "They found that they can get better margins by selling tools. So OpenAccess is opening a lot of doors."
Professor Alain Greiner of the Pierre & Marie Curie University in Paris said that, although his current work is based on a chip-design database developed by the university with Bull Computer, he is looking at how that can interoperate with OpenAccess. "We think this provides an opportunity for universities to re-enter this field and develop new tools.
The chipmakers have yet to find the ideal database for their needs but they are warming to an idea that could streamline an increasingly expensive process and, where it comes to databases, bring the EDA industry into line with the rest of the computing world.