Tiny Transistors And Cold Pizza
On Mar. 22, Intel Corp. will officially launch what is unquestionably the most complex microprocessor ever built--and possibly the most important chip in the company's history. Called Pentium, it packs 3.1 million transistors--incredible even for a chip four times bigger than Intel's current champ, the 486. Not only will Pentium take personal computers to new heights, but it also should help Intel distance itself from increasingly successful cloners--at least for a while. And it's likely to blunt the RISC (reduced instruction-set computer) challenges from the likes of Digital Equipment Corp. and MIPS Computer Systems Inc.
Intel promises that Pentium will crunch numbers like a mainframe, chewing through 100 million instructions per second, or five times the speed of the 486. If Pentium delivers such a leap in power, it should be Intel's cash cow well into the late 1990s, preserving the $5 billion company's status as the world's top chipmaker. If it doesn't, the Midas of microprocessors could finally be dethroned, ending an era.
Pentium looks so good on paper that many industry pundits believe Intel will pull it off. One important sign: Intel felt secure enough about Pentium to delay introduction by six months, so as not to dampen prematurely sales of 486 chips and PCs based on them. "With enough ingenuity, money, and time," says Michael Slater, president of consultant MicroDesign Resources, Intel has found ways to compete with the RISC chips that have already pushed workstation performance into mainframe turf.
MANY MASTERS. Whatever the outcome, just producing the Pentium is a feat. Engineering anything with 3 million components is a monumental challenge. But Intel had to work with transistors so tiny that it would take 500 of them to circle a human hair. And that's just the physics. There were also design imperatives: Come up with a design that comes close to the zippy RISC chips, yet still runs all the software written for Intel's previous microprocessors. Thanks to that software leverage, Pentium doesn't have to match the fastest RISC speeds. Even 50% faster will probably deter not induce users to switch, declares Slater.
Pentium was midwived by a small army of engineers, programmers, and support personnel who struggled for three years to satisfy those requirements and more. This is their story--one snapshot of the treadmill that high-tech companies must keep forever spinning if they're not to lose ground. Indeed, some Pentium designers are now working on its successor's successor--the 786 generation. But first will come the 686, already in research.
The tale of Pentium began in June, 1989, barely two months after Intel's 486 had been introduced. Vinod K. Dham, general manager of the Workgroup Computing Div. and the man who had managed the final stages of the 486 project, began rounding up engineers. It was christened P5, the fifth in a series dating back to 1979.
This was not to be simply a follow-on. Intel had been tracking RISC technology since the early 1980s, and in 1989 it introduced its own RISC chips. Intel's customers and partners were pressing the company to use more of that technology to goose the performance of its traditional chips. Microsoft Corp. wanted a chip fast enough for its ambitious Windows NT operating system. Compaq Computer Corp. wanted workstation-like performance. In 1990, both companies turned to MIPS' RISC design.
Inside Intel, the battle over RISC raged. Many executives pointed out that while RISC chips--including Intel's--run faster than conventional CISC (complex instruction-set computer) chips, they can be slower when using the enormous library of existing PC software. Those programs represent Intel's biggest advantage in the market, they argued, and should not be abandoned. "None of the other guys [the RISC camp] has any continuity," noted Albert Y. C. Yu, the vice-president to whom Dham reports.
Dham's team set out to prove it could build a chip that runs existing PC software and still delivers near-RISC performance. After a few months of brainstorming and soliciting the views of customers--a first for Intel--the concept emerged. The P5 was to be "RISCy" CISC--not only a tongue-twister but also a hideously complex amalgam of technologies. RISC's speed stems from both fewer and shorter instructions. There's less choice, and short instructions are completed faster. Certain RISC chips allow two or more instructions to execute at once. But CISC instructions, which are needed for Pentium to remain compatible with its predecessors, vary in size--and bigger ones take longer to complete. So the key would be a savvy traffic cop to keep the mixed-together operations synchronized.
By early 1990, Dham's team had roughed out a design. Part of the group set out to develop software to simulate the chip. That summer, after months of tinkering, the simulated chip executed its first instruction: It added two numbers. That alone brought a sigh of relief from Intel's executive suite, and the company quickly decided to commit the vast majority of its resources to the Pentium and its successors. As a result of these studies, says Yu, "we really see no limit on future capability."
When 1991 rolled around, Intel rolled out the first "floor plan" partitioning the silicon into various segments, called "rev zero." By spring, the designers were also divided into teams to create different types of circuits--the memory module, the mathematics unit, and so forth.
That's when the real estate squabbles began. The engineers became obsessed with getting their share of the 3.1 million transistors that Intel's production arm promised to cram onto the chip. Frequently, an engineer would barge into a colleague's office, pleading for a few more transistors so his circuits could do what they were supposed to do. As workdays lengthened into the night and weekends vanished, tensions mounted. Once, when Dham heard several engineers arguing over how to reduce the transistor count in a particular circuit, he told them to take a vacation. They agreed to stay home that Sunday.
PAX PEPPERONI. More strains erupted as different circuits were patched together. When some particular operation failed on the simulators, one design crew often blamed another group's work. So Dham appointed a team to mediate and pinpoint responsibility. To ease the friction, design milestones were celebrated with pizza parties, beer bashes, and dinners out. Once, Dham piled everybody into a bus for a party at the Academy of Sciences in San Francisco.
By late 1991, the simulated circuits were doing more than simple additions. Intel transferred the design to hardware emulators--computers with special circuit cards to ape the P5--and began to tune the chip for software such as Lotus Development Corp.'s 1-2-3. Intel was particularly concerned with Microsoft's Windows NT. Due out this spring, it's designed to wring maximum performance from RISC chips. "I wanted to make sure the Pentium architecture was fine-tuned to take advantage of NT," Dham says. These software demos helped turn skeptics. In 1992, Compaq abandoned its plan to use the MIPS RISC design and will use Pentium.
Last spring, the design finally was ready for silicon. The digital plots of 3 million transistors and myriad connecting wires flashed from workstations over a network to systems that produce the negatives, or masks, that are used to print the design on silicon. Three days later the masks were done, and the engineers returned from a meeting in another building to find their offices filled with balloons. Pizza time again.
BUG HUNTERS. On another floor, it was nail-biting time as engineers waited to test the first chips. Behind locked doors, two-dozen people gathered around a special circuit board, and the chip was plugged in. After some prechecks and simulation work, the chip "booted up" in 10 minutes. The same procedure took five days with the 486, says Systems Assurance Manager Michael J. Szeles. Still, there are always some bugs, so the chip was tested for 48 hours straight. A couple of engineers babysat the whole time.
For the rest of the year, the chip was run through batteries of tests. Starting last summer, Intel began shipping Pentium chips to computer makers so they could hunt for bugs. Fewer than 10% of the glitches that turned up were found by customers, compared with about 40% for the 486. Another concern was to help PC makers figure out ways to cool the chip, which runs three to five times hotter than the 486.
How good is Pentium? Perhaps not quite good enough to squash the RISC challenge. Still, says Slater of MicroDesign, the chip does have what's needed for "Intel to continue to hold on to the major share of the market." That's should keep Midas on top.