01 May 1993
From New Scientist Print Edition. To view it in the New Scientist web archive, follow the link below.
Ken Shoemaker was a senior engineer on the design team which worked on the Pentium, the latest chip from Intel, the US's largest semiconductor manufacturer, based in Santa Clara, California. Released this month and costing up to $1000 for a square of silicon just 16 millimetres by 17 millimetres, each Pentium represents probably the most expensive real estate in the world: a million of them would just about cover the floor of a large flat, and cost a billion dollars. If Intel has its plans right, the Pentium will sell enough to carpet a terrace of houses. But rivals think that a combination of delays, unexceptional performance, cooling problems, an unusual design and the inherent difficulties in reliably producing a chip with 3.1 million flawless transistors may show that Intel has bitten off more than it can chew.
Designed for a world where multimedia and video processing will be commonplace, the Pentium is the most complex microprocessor ever developed. Its success or failure could also determine the future of the PC industry, either opening up new vistas of processing power - or, if it fails commercially, putting a ceiling on the capabilities of the millions of PCs already sold because they will never be able to perform the tasks possible on dedicated graphics and scientific workstations.
Most personal computers today are based on an Intel design idea known as the x86 series, which started with the 8086, released in 1978. A less powerful version, the 8088, became the basis of IBM's original PC in 1981; Compaq then built its first PC around the 8086 in 1984. This set a de facto standard for personal computer processors which was widely adopted, except by Apple Computer which stuck with incompatible Motorola processors. Intel followed the 8086 and 8088 with the increasingly powerful and flexible 80286, 80386 and 80486 chips (seeGraph), commonly referred to by the shorthand of 286, 386 and 486. The secret of their success was that software which ran on an earlier processor would run on a later one, but more quickly. Intel claims its processors run in 70 million of the 100 million PCs sold to date.
The Pentium is the 486's successor. (It is not called the 80586 because last year when Intel was suing a rival manufacturer, AMD, for selling chips called 80386s, a judge ruled that unlike a name, a number cannot be copyrighted or trademarked.) It churns through calculations at up to 112 million instructions per second (mips), compared to 54 mips for the fastest 486 chip. Intel claims it will run software between five and ten times faster than the 486 - depending on what software, and which version of the 486.
But, unlike the 486 or any of its earlier siblings, the Pentium will compete against more powerful chips from rival manu-facturers, which are already targetting Intel's markets. Most notable of these are Digital's Alpha, the Sun Microsystems SuperSparc and the R4400 from Mips Technologies, a subsidiary of the graphics computing specialist Silicon Graphics. The Alpha, for example, is claimed to have a top speed of 200 mips.
Intel has tried to stage-manage events by keeping the Pentium under wraps for three-and-a-half years, and gagging its customers - computer manufacturers including IBM, NCR and Hewlett-Packard. However, information seeped out: rivals claim that they obtained inside information through unhappy Intel customers, and have not been slow to point out the new chip's potential shortfalls.
Their main criticism is that the Pentium is being launched a year late. Intel had promised it for May 1992, as confirmed by company executives as late as January last year. Some rivals cite this as indicating that the design process had more problems than expected - or that Intel was struggling after taking on too much.
Intel's own story of the Pentium's evolution is one of long days, late nights, intense arguments and a lot of creative tension. The engineers who worked on the design believed that it would be a major advance in silicon technology, and were dedicated to solving any problems that got in their way.
Shoemaker was involved from the beginning, writing the Pentium's 'logic model' - a description, in software, of its intended functions. A key decision was taken during early brainstorming sessions in the latter half of 1989 when it became obvious to the design engineers, who had spent a lot of time listening to Intel customers, that the corporation had to come up with a chip that could run existing PC software to keep its large installed base happy - but also be seen to be moving in the direction of Alpha, SuperSparc, and the R4400, which by then had been announced though not released.
These rival chips differ from Intel's x86 family in a common feature: they are built to a design pattern known as Risc, for reduced instruction set computing. The 'instruction set' of a processor is the number of different operations it can carry out on pieces of data - such as move it from one memory register to another, test whether it is even or odd, or add it to the contents of another memory register. Performing millions of such simple calculations every second produces the complex responses of, say, a word processing program or a spreadsheet and the screen display and outputs.
Processor designers used to try to include ever more wide-ranging instruction sets so that every possible operation could be encoded with one command, rather than breaking it into component parts. Thus, doubling a number and then adding it to another might be described by one complex instruction, rather than the three simpler ones that it entails.
However, complex instruction sets carry a double penalty: it takes longer for the processor to search through the set to find the instruction to carry out, and more complex instructions may take longer to perform than the total of the simple ones of which they consist. The alternative - reducing the instruction set to the frequently used operations, and splitting complex instructions into their simpler components - means a processor can work at high speeds. It also reduces the number of components such as transistors needed to form the logic gates that perform the operations. The Pentium is not a Risc chip; its rivals have half as many transistors - just 1.4 million on the Mips R4400, and 1.67 million on Digital's Alpha.
Until that 1989 brainstorming session Intel had been determined not to use Risc at all, and to stick with complex instruction set computing (Cisc). As Shoemaker says, 'The Pentium had to run software that would run on a 386 PC, so the team had to ensure that it would cater for the needs of the existing software architecture. We had to be certain that Pentium would reach the same result in any calculation as a 386 or 486 would have done.' But at the same time, its potential power would bring it into direct competition with the chips running graphics and scientific workstations, a fast-growing market sector where Risc is popular because the work consists of repetitive, intensive number-crunching. If the Pentium ignored Risc and could not compete in performance terms, it would lose out. But if it embraced Risc, it would not be compatible with its family.
These conflicting pressures produced heated debate within the company, and considerable tension during later stages of the design process, because the result - a compromise between old and new which Intel has dubbed 'Crisc' - created extraordinary difficulties for the design team. Shoemaker points out that not many engineers have ever faced such a complex design challenge: to produce a chip that can perform well with old software, and do something new. The design team on Alpha, for example, had a clean slate. 'When you come out with a brand new design, even mistakes can be features,' Shoemaker says. 'But with a compatible chip, a mistake is a mistake.'
Decisions on how to implement the Crisc concept were taken in those early months. One of the most critical was the choice of a 'superscalar' architecture, a key feature of many Risc processors, to handle the integer calculations which form the bulk of work done by general-purpose computers. Superscalar architecture means the chip can simultaneously execute two sets of instructions on two different pieces of data. Ideally, it could do twice as much work in the same amount of time. In reality, it does not quite achieve that because sometimes one calculation is waiting for data being calculated in parallel - rather in the way that widening a one-lane highway to two lanes will not necessarily double the traffic flow (because some vehicles change lanes), but will boost it considerably.
Other key Risc-type techniques that were incorporated into the Pentium's design were floating point arithmetic, branch prediction techniques and on-chip memory caches. Floating point arithmetic is used in Risc processors because it is believed to be the best technique for heavy-duty calculations such as complex mathematical and scientific data. The 486 processor has a floating point capability which has been overhauled for the Pentium: Intel claims to have tripled its performance.
Branch prediction techniques increase processing efficiency by allowing the hardware to 'guess' ahead at instructions that are coming to it, and prepare for them. The memory caches on the chip store the most frequently used instructions, to cut the time spent searching for them.
FORMAL PROOF OF THE PUDDING
The preliminary design of the Pentium, in the form of the software model describing how its component mathematical logic gates would interact, passed its first milestone in the summer of 1990 when it was run on the computer-based simulation program that Intel uses in all its chip design, and it successfully added two numbers together. For the first time, the design team knew that the Crisc approach could work.
It then took about a year to turn the mathematical logic model into a design layout of electronic circuits on silicon. The team - by now consisting of several dozen engineers, aged between 25 and 45 - split into small groups, each responsible for a different part of the chip. They worked 14-hour days, ignoring weekends and holidays. Discussions took place on biking trips and in corridors. The pressure was relieved by occasional parties and games of 'Kooshball' - a Silicon Valley invention involving a ball made out of elastic bands. But it was over integrating their ideas onto one chip that the real arguments began.
The team had been told at the outset of the project that the chip could not contain more than three million transistors, for a number of reasons. First, the larger a chip is, the longer the circuit paths and so the less efficiently it works. Secondly, more transistors consume more power, which can cause problems as the circuits heat up. Then, the more circuit elements there are, the greater the chance of fabrication errors - and just one can render a chip useless. And finally, a larger, more complex chip will be more expensive, and Intel had to hit the right price.
Although three million transistors should have been enough to include everything, the team members found that it was not; they had to make compromises. The discussions about what could be cut and what could stay took hours. 'Everyone had staked out the probable area they'd need on the chip. Some estimates were high and some were low,' recalls Shoemaker. 'Those who could easily fit into their area didn't want to give it up until they absolutely knew they didn't need it. But ultimately, everyone had to work together.'
By late 1991, the Pentium was ready for more extensive simulation. Well before their creation was put onto silicon, the designers needed to ensure that it would work in real life as it was meant to. The logic model was enhanced to include all the changes: 'It was a continual process of refining it to turn it into the full-blown architecture,' says Shoemaker. Just to make sure, it was run through millions of clock cycles on the simulator.
In April 1992, although further refinements could still have been made, the team felt confident enough to transfer the logical model to silicon - which prompted Shoemaker's moment of epiphany. Pentiums were then beta-tested using computer manufacturers - to let them discover any further bugs. Intel claims it was this extensive testing and debugging that held it up. 'We had hundreds of thousands of tests,' says Ron Curry, product line manager for the Pentium, 'largely as a result of meetings with our customers and their requirements that we do super-heroic tests before we run with a product.'
RUNNING AGAINST THE CLOCK
But after all that, has Intel got it right? The Pentium may look brilliant against the 486, but compared with the leading Risc processors, it is somewhere in the middle of the pack.
The key benchmark for microprocessor chips is the SPECmark, obtained by running 20 standard computer programs, timing them and calculating their geometric mean. Two separate SPECmarks are obtained for integer and floating point performance; they are also measured at different clock speeds (the frequency at which the processor operates).
At a clock speed of 66 megahertz, the Pentium has an integer SPECmark of 64.7 and a floating point SPECmark of 56.9. This puts it way behind the Alpha, which at 133 megahertz has an integer SPECmark of 74.8 and a floating point SPECmark of 112.5. (The Alpha can also operate at 150 megahertz and 200 megahertz.) It also trails the latest Mips processor, the R4400, which at 75 megahertz has an integer SPECmark of 84.2 and a floating point SPECmark of 104.2. Mips's earlier processor, the 50 megahertz R4000, was closer to the Pentium: it had an integer SPECmark of 58.3 and a floating point SPECmark of 61.5.
To this, Curry says that Intel intends to manufacture in volume, and does not intend the Pentium to be a top-of-the-range processor. He argues that Intel sells its chips in millions, whereas Digital or Mips will only sell in 'niche market' quantities - of tens of thousands. Clearly, Digital and Mips will have their sights set higher.
Not surprisingly, Mips is the Pentium's harshest critic because it is keen to muscle in on the mass market for desktop computers that Intel dominates. Andy Keane, Mips product marketing manager for microprocessors, points out that the Pentium's power consumption means it will get considerably hotter than the 486, so PC manufacturers could have difficulty putting it inside a compact box thatsits on a desk because they would not be able to generate sufficient air flow to cool it. In that case the chip could over-heat and thermoelectric effects, rather than semi-conduction, take over. The result: the chip would make mistakes.
Too hot to handle?
A chip's power consumption depends on, among other things, its clock speed and operating voltage. Intel says the Pentium consumes 13 watts. Keane calculates that for a 66 megahertz version it would be closer to 17 watts, compared to the highest specification 486, which uses just 6 watts.
The Pentium's rivals also heat up noticeably: the Alpha can consume anything between 15 and 30 watts, depending on the clock frequency; the R4400, up to 17 watts. But Keane argues that they also offer twice the processing power.
Martin Lomas, European programme manager for the Alpha, explains that machines based on the Alpha have a cooling tower, but that this is normal for a high performance chip. Cooling is a much greater problem in smaller machines which have no fans, such as laptops and notebooks; but because they are the fastest-growing sector of the computer market, Digital is designing a low-power version of the Alpha, with a lower clock speed and voltage, and possibly with fewer features.
To ever get the Pentium into notebooks, Intel will probably have to take the same route, although it is not saying so at the moment. Most manufacturers intending to make Pentium-based machines appear to be making only network servers, which direct data traffic within a larger network of PCs. These would not need to sit on a desktop, but will make the most of the increased processing power.
The final question is whether manufacturers will get as many Pentiums as they want. Keane says that the huge number of transistors on the Pentium will result in a low production yield. 'That whole little square on the silicon wafer has to work. If there's one transistor that doesn't work, you have to throw (the whole chip) away,' he says. Risc designs like the R4400 and Alpha, having half as many transistors, can in theory have much higher yields than complex designs.
But Intel is stuck - perhaps happily - with its complex design. Now, recalling the effort put into the years of design, Shoemaker says 'you'd go home and you'd be detached all night worrying about a particular thing. You always have a fear that you're not going to make it, but you have faith that you are.
Monica Horten is a freelance journalist specialising in information technology.