Such flaws made the 1103 a spectacularly stubborn and perverse contrivance. Its patriarch, Gordon Moore, termed it "the most difficult- to-use semiconductor ever created by man." It was also hard to manufacture. Intel had so much difficulty turning out an economical volume of working chips that it had to assign entire teams of engineers and technicians to the drudgery of picking good chips out from the river of useless silicon coming off the fabrication line, a job so fervently detested it was labeled "turd polishing."
To the CSL team, however, the 1103 s shortcomings were obstacles to be overcome. "It seemed pretty clear to us that the memory should be semiconductor, although we didn't really know whether those chips worked—and it turned out later that they don't," Lampson recalled. "We certainly would have preferred a more robust chip. But we were very confident that by putting in error correction we could make up a very satisfactory system that would work." That was an understatement. As CSL well knew, if they could only overcome the 1103's manifold obstinacies, their machine would boast the speediest and most reliable memory on Earth. CSL's resolution of the PDP-Sigma imbroglio failed to quell entirely SDS's discontent with the outcome. From El Segundo was heard continuous carping that PARC had caused the division insupportable embarrassment by spurning its top-of-the-line product. This provoked Pake into an outburst that settled the matter once and for all. In a blistering memo he reminded headquarters that his best new engineers had voluntarily agreed to suspend bona fide research projects for the year or more it would take them to satisfy SDS's querulous concerns. "It is unthinkable to me that Xerox sets me the task of hiring creative, imaginative, top-rank researchers and then expects me to insist that they handcuff themselves with inappropriate equipment," he wrote. "I will do my best to provide them with the kind of first-rate technical support it is reasonable to expect in Xerox research laboratories. If that is the wrong way to build a first-rate corporate research center for Xerox, then I am the wrong man for the job."
Meanwhile, the CSL rank and files lingering resentment at their Southern California in-laws was manifested when the time came to give their clone a name. One afternoon a group of engineers gathered at someone's house to mark an intermediate milestone on the project.
"A moderate amount of beer had been consumed and we were trying to figure out what to name this thing and there was considerable hilarity," Thacker recalled of the day they came up with the formal moniker of "Multiple Access Xerox Computer." It sounded conventional, but everyone on the scene got the joke. In honor of the man who had sold his lousy computer company to Xerox, the first major project undertaken by PARC would be known for all time by the acronym "MAXC." No member of the lab ever forgot to remind outsiders, "The 'C' is silent."
MAXC's christening was more than an opportunity to tweak Max Palevsky.* It crystallized their awareness that what was coming together in the Computer Science Lab was no longer a DEC machine, but their own.
"We did everything, from soup to nuts," McCreight said. Liberated from slavish adherence to the PDP-10 design, they were able to get dozens of functions running faster or cheaper. McCreight performed one such feat with his disk controller. Conventional disk controllers were generally equipped with their own separate processing units, like Stegosauruses with their second brains, which added significantly to their cost and complexity. Poring over the system schematics in his office one day, he was struck by the realization that there would occur certain periods when, having executed one instruction and not yet received the next, MAXC's central processor would be idle but available, like a car left running unattended in the driveway.
*The point was probably lost on the target. As Pake said later, "I doubt that Max Palevsky ever cared, or even knew about it."
"I learned enough about the processor to realize I could use some of those spare cycles," he recounted. "In effect I could kidnap the processor to do some arithmetic for the disk controller. I wouldn't have to put so many gates into the disk controller"—saving another few thousand dollars in hardware—"if I could periodically borrow the processor to compute some of the things I needed to compute." McCreight's realization was their first embrace of the concept of "multitasking"—giving the processor numerous jobs to juggle at once. Implemented on this modest scale in MAXC, it was destined to pay enormous dividends later.
Meanwhile, they continued to inject refinements into the PDP-10 design. If along the way they discovered some flaw, an inelegance or vulgarity committed by the original designers, they had no compunction about fixing it. At least once this resulted in making MAXC too good. This was the episode of Fiala's floating-point bug.
Floating point operations allow computers to handle huge numbers by breaking them into two pieces: the mantissa, which comprises the significant digits, and the exponent, which is a power of ten. Thus the number 632,100,000 would be split into a mantissa of 6.321 and an exponent of eight (i.e., 10 to the 8th power). To multiply two numbers in a floating-point operation, the computer simply multiplies the mantissas and sums the exponents.
Floating point functions are critical to the efficient use of a computer's resources. But they are among the most difficult to properly implement in hardware. Sure enough, while coding the floating point microcode Fiala discovered a number of bugs in the PDP hardware, or at least places where he could improve on it. Without thinking twice, he did so. As McCreight recalled, "He figured it's his floating point, and it was up to him to make it more accurate than the PDP's."
One of the classic frustrations of systems design is that fixing a bug in one place often creates others elsewhere, the way squeezing a balloon at one end makes it bulge out at the other. Something like that happened in this case. When the MAXC team tried to launch a program called Interlisp that had been written for the original PDP-10, "for some reason we couldn't bring it up," McCreight said. "We couldn't even find nil, which is the first thing you look for in Lisp—if you can't find nil, you're in trouble." The problem, as they discovered after hours of tedious investigation, was that Interlisp utilized the PDP's inefficient floating-point algorithms to execute its own code. Fiala's fixes left the program hopelessly confused, as though someone had rearranged its furniture in the dark. Unfortunately one reason they were building MAXC was to run Lisp. So with a great show of reluctance, Fiala acceded to pressure from the rest of the lab— and programmed the bugs back in.
As they anticipated, Intel's 1103 memory chips proved to be a major nuisance. The reputedly serviceable chips that survived the "turd polishing" stage still arrived at PARC with their deeply flawed design intact. Among the headaches was their so-called pattern sensitivity: Certain combinations of bits would cause an intermediate bit to "flip," so that a sequence of 1001 might read out incorrectly as 1101 or 1011.
Thacker overcame the fault with an error correction system that could identify and reflip erroneous bits. This worked as long as they only occurred one at a time. Eventually, however, a second bit would fail. "Then," Thacker recalled, "you'd get an honest error and the system would crash and we'd have to change the chip." That was not as rare an occurrence as it might seem. For a time MAXC boasted the largest semiconductor memory of any computer in the world, an achievement that temporarily made PARC Intel's single biggest customer. Approximately 25,000 of the 1103s got packed into four cabinets standing six feet high, each one containing four card cages with sixteen circuit boards that in turn were each about the height and width of a standard sheet of typewriter paper and held ninety-six chips apiece.
The memory boards were the only part of MAXC the lab sent outside for fabrication. "There were so many of them—two hundred and fifty-six, not including spares—that it was economical to make a printed circuit board for the memory, which was not an inconsiderable task in those days," Thacker recalled.
At one point the need for a dependable board-maker threatened to set Thacker on a life of crime. Intel had sublicensed the manufacture of 1103s to a Canadian company called Microsystems Internation
al Ltd., or MIL, a subsidiary of the Canadian telephone company. "They offered to make boards at a substantially lower price than we were paying," Thacker said. "I remember going to Ottawa one time carrying a sample memory board with me, at that time a one-thousand-dollar object."
As he breezed through Canadian Customs, he was stopped by an officer who asked what he was carrying.
"It's a printed circuit board," Thacker replied.
"You'll have to pay duty on that."
"But it's just a prototype. I'm bringing it in today and bringing it back out tomorrow."
"In that case you'll have to pay duty both ways. What's it worth?"
Thacker thought quickly. "About fifteen dollars."
"Oh," the officer said, waving him through. "If that's all, don't worry about it."
Notwithstanding the stubborn 1103s, MAXC proved a superbly robust machine thanks to Thacker's inspired design and the lab's resourceful craftsmanship. Because Thacker had designed in more physical memory than was needed for all its logical operations, whole memory boards could be pulled out for repair without taking the system down for even a nanosecond. Once fully debugged, the machine set records for uninterrupted availability on the ARPANET, handily outperforming computers that had taken squadrons of engineers years to design. In contrast, the Computer Science Lab at PARC implemented MAXC in scarcely more than eighteen months. The cost to Xerox was about $750,000, of which roughly a third went for the memory.
All this was accomplished under intense deadline pressure. "Everybody was waiting for MAXC to exist," recalled McCreight, whose office at Porter Drive was so crammed with equipment—six-foot-high racks holding twenty-four spinning disks on two spindles, arranged one on top and one below like pizza ovens—there was barely room for human beings. "Every day guys would show up in my office and say, 'How soon, is it coming now?' So we knew it was a very much desired thing. I mean, the point of the lab was to program and we couldn't program until we had the computer system."
As Taylor anticipated, moreover, the seemingly make-work project paid exponential dividends in group dynamics. At first glance, devoting so much time and money to reproducing a computer available on the open market seemed sheer profligacy. But from Taylor’s point of view, the assignment to produce a real machine had given his engineers a unique opportunity to parse out their own strengths and weaknesses in ways Taylor could never have devised himself. What emerged at the end of the program was a seamless, remarkably powerful unit.
"In a small group the dynamics are like those on a good basketball team," Kay observed. "Everybody has to be able to play the whole game. Each person should have certain things they're better at than the others, but everyone should be pretty good at everything." MAXC proved they were all pretty good at everything that mattered: hardware, software, microcoding, programming. "They made having to do MAXC into a virtue. No matter how you slice it, the job was amazing. It was not trivial. Not even close to trivial."
It was, however, merely the first step in a long journey. MAXC was barely finished before they started thinking about what to do next.
CHAPTER 8
The Future Invented
While his lab staff occupied themselves with implementing MAXC in the spring of 1971, Taylor got around to one of the keystone tasks Pake had set down for him: He recruited his own boss.
This did not happen a moment too soon. Taylor was in critical need of a buffer between himself and the rest of the organization. He and Pake were still on speaking terms, but there was little more to say about their relationship. "It was like a bad marriage where two people stay together because of the kids," remarked Rick Jones.
No one would dispute that "the kids" were worth the effort. CSL's work on MAXC provided just a hint of what they might be capable of once they hit their stride. But accepting Taylors crew at this level involved a sort of Manichean bargain: Perhaps the intellectual tension the Computer Science Lab generated on Porter Drive helped goad the other labs into matching their energy and creativity, but they also made things harder than was necessary. Pake, who had literally put his job on the line for MAXC, observed privately that his efforts to build a bridge between PARC and SDS might actually have borne fruit if only the CSL engineers had not taken every opportunity to belittle El Segundos work. Pake's proclivity for solving conflicts by splitting them down the middle and Taylors for establishing and holding his position in the local pecking order stood in direct opposition to each other. This situation was not destined to improve.
At least Taylor had the wisdom to see that his buffer to the outside world should in training and experience resemble Pake more than himself. The man in question was yet another charter member of the ARPANET clan. In the late 1960s Jerome I. Elkind had been in charge of computer research at Bolt, Beranek & Newman (BBN), a firm of Boston engineering consultants known familiarly as BBN. There he had supervised the firm’s successful bid to build a critical piece of Taylors cherished nationwide network. This was the system of "IMPs," or Interface Message Processors. The IMPs formed a subnetwork of standardized computers (they were remodeled Honeywell minicomputers) that stood as a gateway between each host s mainframe and the central network. In effect they functioned as universal translators, allowing the network to interconnect dozens of incompatible computers without turning into a cacophonous Babel. (The concept, which solved one of the fundamental technical problems bedeviling the ARPANET'S designers, was the brainchild of Wes Clark.)
Thanks mostly to its extensive work on the ARPANET, Bolt, Beranek & Newman became one of ARPA's largest private contractors. This circumstance had forged an amicable relationship between Elkind and Taylor. Elkind also had bonds with several other PARC people, including Butler Lampson, who he had met when his BBN division bought one of the first SDS 940s, and Peter Deutsch, who had worked at BBN as a high school student and during summer vacations from Berkeley.
Elkind was an empirical-minded scientist with a conspicuous streak of skepticism. This temperament elicited sharply divergent reactions from his peers. Some appreciated his discretion—a trait which, after all, would not be such a drawback for the manager of a lab venturing to the edge of the unknown. Others found him an insufferable pessimist whose disposition was certain to clash sooner or later with Taylor's enthusiastic cajolery. Even at BBN, observed Severo Ornstein, "Jerry was not universally liked as a technical supervisor. I think he didn't have the right touch."
Taylor reassured the lab, however, that his and Elkind’s personalities would be complementary, not contentious. He saw Elland as playing "Mr. Outside" to his own "Mr. Inside"—as a sort of human IMP providing a painless interface between CSL and PARC. Elkind could handle the bureaucratic rubbish for which Taylor had no patience, leaving him free to keep to his role of evangelist, guru, and all-around father figure. Presented so abstractly, the arrangement almost seemed rational.
Taylor laid the groundwork carefully for Elkind’s recruitment. He asked Wes Clark, an Elkind admirer from his MIT days, to pass on a glowing recommendation to Pake, and called Elkind himself to sell him on his Outside/Inside plan before he met with Pake. Elkind listened, intrigued, but his understanding of the arrangement never fully matched Taylor's. "I always thought Bob's role was to be there as a very strong associate director," he said later. "My role was that I was going to be managing the lab."
It was not that Elkind was intent on micromanaging his researchers. On the contrary, as a research chief he had always believed in granting his best people a large measure of independence. "The style of research that I had been used to at BBN was certainly one of very strong principal investigators ordinarily doing work on their own," he recollected. But he also viewed it as the managers responsibility to impose a group philosophy—"a vector in certain areas" so they would not be "proceeding off at random." As for the proper vector of research at PARC, "the fact that we were a part of Xerox meant that one would spend a great amount of time and effort doing things that were useful t
o the corporation." If upon hearing these words Taylor felt any impulse to tell Elkind to leave the vectoring to him, he stifled it—for the moment.
One glorious spring day George Pake made his way to Elkind’s house in a quiet suburb west of Boston to take his measure. They had never met, but everything Pake had heard about Elkind, including Wes Clark's fulsome praise, predisposed him to like the man. Over the course of a few gratifying hours under the crisp New England sun he satisfied himself that Elkind was everything Taylor was not: a sober scientist with indisputably sober credentials, among them an MIT engineering doctorate (achieved under Licklider, no less). Elkind's research experience was unassailable, as was his experience managing large research teams in a corporate setting. His gravity did not for a minute faze Pake, who interpreted it as seriousness of purpose. Jerry Elkind was exactly the sort of high-caliber manager he had dreamed about placing in charge of the labs ever since PARC had opened its doors. (Instead he had gotten the reluctant and distracted Bill Gunning and the academically suspect Bob Taylor.)
Dealers of Lightning Page 15