wdv-notes Stand: 29.DEZ.1994 (2.) 329 Intel Pentium – Business Must Learn from the Debacle. Wiss.Datenverarbeitung © 1994–1995 Edited by Karl-Heinz Dittberner FREIE UNIVERSITÄT BERLIN Theo Die kanadische Monatszeitschrift The Im folgenden sowie in den wdv-notes verbreitet werden, wenn dabei die folgen- Computer Post, Winnipeg veröffentlicht Nr. 330 werden diese Artikel im Original den Spielregeln beachtet werden. in mehreren Artikeln in ihrer Januar-Aus- nachgedruckt. Der Dank dafür geht an Permission is hereby granted to copy gabe 1995 eine exzellente erste Zusam- Sylvia Douglas von The Computer Post, this article electronically or in any other menfassung des Debakels um den Defekt 301 – 68 Higgins Avenue, Winnipeg, Ma- form, provided it is reproduced without des Pentium-Mikroprozessors [1–2] des nitoba, Canada, Email: SDouglas@post. alteration, and you credit it to The Compu- Computergiganten Intel. mb.ca. Diese Artikel dürfen auch weiter- ter Post. Intel’s top of the line Pentium™ microproc- This particular error in the Pentium was in essor chip has turned out to have a slight flaw the floating-point divide unit. Intel manage- Editorial: The Computer Post – Jan.95 in its character: when dividing certain rare ment was concerned enough about it that pairs of floating point numbers, it gives the they pulled together a special team to assess The Way it Will Be wrong answer. For anyone who owns a Pen- the implications. tium-based computer, or was thinking about This month you’ll find we’re reporting a In the words of Andrew Grove, Intel’s CEO, buying one, or is just feeling curious, here lot of background information on Intel’s posting later to the Internet [3], “We were are... Pentium problems. In fact, we have never puzzled as to why neither we nor anyone else before devoted this much space to any sin- The Facts about Pentium had encountered this earlier. We started a gle news story. separate project, including mathematicians We’re not just picking on Intel because By SYLVIA DOUGLAS and scientists who work for us in areas other they’re big – although there is a certain The Computer Post than the Pentium processor group to exam- fascination in watching such a large corpo- ine the nature of the problem and its impact.” ration being forced by public pressure to Like an avalanche, it started small. One Inexplicably, though, the now-infamous reverse a strongly held position. What’s little whisper that something was wrong. Sud- FDIV bug didn’t make its way onto any lists of new and different about this drama, what denly, without any more warning than that, known bugs. Intel didn’t inform even their deserves serious study by everyone running everything was sliding out of control. Giant largest customers. a business in the second half of this decade, Intel, the company that makes chips for 80% is the role played by online communica- of the world’s PCs, found itself in the middle Instead, the special team did their analysis. tions. of the worst PR debacle in its corporate In Grove’s words, “This group concluded Let’s not exaggerate. The Internet didn’t history. after months of work that (1) an error is only do it all. Intel didn’t give in until the Pen- likely to occur at a frequency of the order of tium story became widespread enough to Five years from now, one way or another, once in nine billion random floating point this will be an excellent case study for busi- affect stock prices and threaten future sales. divides, and that (2) this many divides in all This still means CNN and USA Today, not a ness schools in their crisis management cour- the programs they evaluated (which included ses – as an example of how not to do it. It was newsgroup and a few Web sites – not yet. many scientific programs) would require At the rate people are connecting to the a minor technical problem, Intel management elapsed times of use that would be longer thought, just like many others they have dealt Internet now, within a few years the public than the mean time to failure of the physical at large is going to be looking online first for with over the years. Last June, just when Intel computer subsystems.” was in the midst of ramping up its Pentium their news and information. More than that, production to meet the challenge of the online will be where they can ask their own PowerPC chip (coming from Apple, IBM, and While those months passed by, many more questions, express their own opinions, and Motorola), some of its engineers noticed a scientists and engineers, persuaded by In- discover what other people think. So it is wrong answer in one of the multitude of tests tel’s advertising that the Pentium was now vital for business to understand the new they were running on the Pentium. able to give workstation performance at af- dynamics of an interactive customer base. fordable prices, adopted this platform for criti- Here is a case study of how it happened to cal calculations. The price of Pentiums started one company. »You know postmodern – to come down, and businesses adopted the Intel ran into the phenomenon early be- platform for calculation-intensive spreadsheet cause of who they are, and who their cus- Well, this is PostReality.« work. After all, one of the advertised advan- tomers are. They sold the Pentium to scien- The Computer Post, Winnipeg. tages of the Pentium was its very fast floating tists and engineers who had been connected point arithmetic. to the Internet for some time and were comfortable with expressing themselves that All chips have errors. This is a fact of life in In what now turns out to be Intel’s one bit of way. the semiconductor business. Connecting mill- good news, the platform was also adopted by Already, a controversy on the Internet is ions of transistors is not something that can large numbers of consumers looking for a very, very public. Many business people are be done error-free the first time. The normal fast, capable multimedia computer to run all now connected, and a rapidly growing practice is to collect up a number of these the exciting new games and educational soft- number of private individuals. There were errors and their design fixes, and apply them ware coming to market. more than ten thousand messages posted to to a revised version of the wafer mask which The report produced by the Intel team has comp.sys.intel, and there is no way of deter- then becomes what the production lines are now been made public. We have included mining how many times that number of churning out. excerpts from this “White Paper” on page 12 people scanned the newsgroup for informa- tion. In software this is called a slipstream up- [Editors Note: See [4] ]; Intel will send a copy to anyone who asks for it. Journalists were certainly among them. grade; most software companies do it. In This past year, only the leading-edge re- hardware, it’s called a stepping. Chips can go Curiously enough, though, it was not Intel porters were using the Internet as an infor- through 5, 10 or even 20 steppings in the that made the first public announcement of mation-gathering tool; two years from now, course of their lifetime. Normal practice is to the bug. Dr. Thomas R. Nicely of Lynchburg they will all be doing it. It is a matter of keep at least major chip customers updated College, a professor of mathematics, was moments to follow up a Web page reference with a current list of outstanding bugs in each doing some work in number theory that in- in a newsgroup article and find a detailed chip. volved taking a lot of reciprocals. To double- source document, or to post a question and check his work, he did it twice – once in To be continued … © 1995 – The Computer Post, Winnipeg floating point, and once in integer arithmetic. Die 1. Edition erschien bereits am 23. Dezember 1994. Seite 1 wdv-notes 329 Back in June, Nicely noticed a small dis- had developed a software workaround that On December 20, Intel caved in to the crepancy between his two sets of results. Of could at least be compiled into programs pressure. They announced that they would course, it did not occur to him at first that this where source code is available [8]. They take an unspecified reserve against their was a problem with the Pentium chip. He didn’t meet each other in person – all the work fourth-quarter earnings, and “exchange the spent a lot of time eliminating other causes, was done via the Internet, and all results were current version of the Pentium processor for such as compiler artifacts and errors in the posted publicly for the world to see, criticize, an updated version, for any owner who re- PCI bus. and contribute to. quests it, free of charge anytime during the life of their computer.” On October 22, after running his test case Meanwhile, the implications of the FDIV on a number of 486 and Pentium computers, bug began to sink in. Vaughan Pratt of Stanford Most of the trouble could have been avoided he became certain the fault was in the Pen- University’s computer science department if, in early summer, Intel had not tried to tium CPU. That day, he contacted Intel tech started examining the Coe model, and did conceal the defect. They might have lost support. After eight days, they had no an- some analysis of Intel’s probability analysis.