Mad Dog 21/21: Two-Minute Warning
September 29, 2014 Hesh Wiener
In the Super Bowl, the warning two minutes before the end of the first half tells the halftime entertainers to prepare for a big cash outlay. The second reminds players they can soon resume personal pursuits, such as domestic violence. IBM’s financial year has warnings, too. In the first quarter IBM warned it was relinquishing its chip factories. By the second it seemed the taker would be Global Foundries. During the third it emerged GloFo wanted IBM to pay a big disposal fee. The fourth will be a nail-biter; vexatious but better than a wife-beater.
The Wall Street Journal claimed, in August, that the NFL had discussed the sale of the right to appear during this year’s halftime show with three pop music acts: Coldplay, Rihanna, and Katy Perry. In this and related reports the media explained that a Super Bowl halftime performance gave a big boost to an entertainer’s subsequent earnings from concert appearances and recordings. Making the performers pay would, it turned out, not to be as dramatic a change as one might think. In prior years, the performers worked for free, and the NFL apparently coughed up some money to cover expenses. The expenses involved were measured in six- or seven-figure numbers. The benefits that subsequently accrued to the entertainers might have run an order of magnitude larger. After all, this is the Super Bowl, the mainframe of sports, where TV ads cost a few million bucks a minute for airtime and lots more in production costs. Ads shown at halftime are likely to get the most viewers because the mid-game break turns out to attract more attention than any other part of the event. Last year’s show attracted over 115 million viewers.
While the sums involved in the Super Bowl halftime activities are substantial, they are tiny compared to the fortunes involved in IBM’s current effort to dispose of its chip factories. If the deal goes down along the lines of what has been reported, up to two billion simoleons more could change hands. That’s a lot more than Katy Perry and her entourage sees for a concert or during a whole concert season for that matter. It’s probably more than she’ll gross in a lifetime, although the way pop superstars run their businesses these days, I wouldn’t want to rule the possibility out. Perry, or her rivals for this year’s Super Bowl halftime gig, Rihanna and Coldplay, could be worth as much as a chip fab. They certainly will last longer and in fact already have. Fabs have to be replaced or totally renovated after several years. Pop stars can generally get away with less maintenance: a little botox, some plastic surgery, maybe a stint in therapy or rehab now and then.
But the cost to IBM of kissing off it chip factories, perhaps $2 billion before the beancounters dress it in mathematical mufti, might actually be a price IBM can bear more easily than the cost of running fabs at a loss that might well be between $1 billion and $2 billion a year, to say nothing of any environmental cleanup or other endgame expenses that might be incurred during an eventual shutdown as fabs age beyond the point of refurbishment or salvage.
When IBM first talked publicly about its plans to get out of the semiconductor fabrication business, it explained its decision by pointing out that the semi fab business was a low margin pursuit and that it had lost interest in low margin endeavors. IBM has been shedding what it considers low margin operations for quite a while, getting out of PCs, disk drive manufacturing, printers, point-of-sale systems and more. It is quitting the low end server business as quickly as it can tie up the loose ends on its planned sale to Lenovo. But Big Blue isn’t on a one-way street. IBM is adding as well as dropping activities. IBM has been bulking up by acquiring endeavors related to cloud information services. And it has been burning the midnight oil trying to monetize the natural language processing and artificial intelligence technology is promotes as its Watson brand.
What makes the chip fabrication business different from other operations IBM has sold off is the way its fruit lies at the heart of IBM’s proprietary server production. IBM’s chip factories make the Power processors used in AIX and IBM i servers and their cousins, the z12 CPU circuits. This vertical integration of IBM’s processor manufacturing makes Big Blue’s products ultra-proprietary. But the proprietary processor products of IBM’s chip factories don’t support the plants and related facilities. In fact, IBM’s CPU chip manufacturing business is just a tiny part of the company’s total semiconductor production.
If they are typical of the plants used by other semiconductor makers, IBM’s chip factories can produce thousands of wafers each month. IBM’s wafers are made in two sizes, 200mm and 300mm. By various estimates, a 300mm wafer can yield about 200 of the physically large CPU chips used in servers. And it makes for a nice round number to use when comparing IBM’s chip-making capacity to its processor chip requirements.
In the case of the mainframe, IBM’s largest machines, when processors are fully populated they use about 120 cores (of which 101 are available directly to users), plus a few spares and systems control engines. The result is a total of up to 4 motherboards, which IBM calls books, each with 6 chips. This works out to 24 of the six-core chips. If IBM achieved 100 percent yield, meaning all the chips on a wafer worked to spec, a 300mm wafer could made eight fully loaded mainframes. If only half the chips work–and it is likely that IBM’s processor chips have a higher yield–than a single wafer would provide all the CPU chips needed for four fully loaded mainframes.
IBM’s mainframe design uses a mix of chips with four, five or six working cores; this improves the chip yield because imperfect chips can still provide all the live cores needed for fully configured computers.
IBM mainframes often have fewer than the maximum four processor books permitted in an EC12 and IBM’s smaller BC12 computers use fewer chips, either two or four per machine. Two-chip machines have nine live engines using a four-working-core and a five-working-core chip. Four-chip machines have two of the nine-live-core processor subassemblies. No BC12 computers give users access to 6 cores on any chip. Thus a single perfect wafer with 200 chips could provide CPU circuitry for up to 50 of the largest BC12 systems and up to 100 of the models with a single nine-core processing module. With a 50 percent yield, a single wafer could furnish silicon for 25 big BC12s or 50 smaller ones.
IBM is believed to ship between 1,000 and 2,500 processors per year. IBM doesn’t disclose the distribution of processors per box but a little back-of-the-envelope guesstimation shows that IBM doesn’t need all that many wafers for a year’s processor production. If IBM built only fully loaded EC12 boxes and sold 2,500 of the systems it would need 625 wafers at a 50 percent chip yield and only 313 wafers if the wafers were all perfect. If IBM shipped only 1,000 systems and they were all small BC12s it would need only 20 wafers at a 50 percent yield or 10 perfect wafers.
Suppose the mix of IBM mainframes sold during one year averaged out to 2000 processors with a mix of configurations that required an average of 20 chips. That would require 200 perfect wafers or 400 wafers with a 50 percent yield.
If IBM’s 300mm fab at Fishkill has about the same throughput as that of Global Foundries in Malta, New York, IBM could make 60,000 wafers a month. In 1 percent of a month, the fab could bake 600 wafers. The creation of 400 wafers at such a facility would then require only two-thirds of 1 percent of a month’s capacity. Because wafer production is a time-consuming process with many steps, the creation of a single wafer typically takes three to four weeks. But if wafer-making were instant, making 400 wafers at a plant churning out 60,000 a month would take only about 5 hours. IBM could make all the chips it needed for a generation of mainframes that lasted two and a half years–recent generations have been shorter, more like two years–in about half a day.
Even if this quick and dirty guesstimate is wrong, even if IBM needed ten times the chip making capacity we’ve reckoned, IBM sure doesn’t need a whole fab to make all the chips its z mainframes require, that’s for sure.
The total Unix market last year was under 30,000 machines. If these machines used an average of four processor chips each and all used IBM’s eight-core Power7+ circuits, total production would require 120,000 chips. At 200 chips per perfect 300mm wafer, this works out to 600 perfect wafers or 1,200 wafers with 50 percent yield. That means all the Power chips needed to make all the Unix (or IBM i) servers in the world, if the machines all used Power and averaged four loaded sockets per machine, would require 2 percent of a single GloFo plant’s monthly capacity. Here, again, it doesn’t look like IBM needs a fab of its own to provide Power processors for the family.
IBM and Global Foundries already exchange quite a bit of knowledge. Chances are that GloFo could make IBM chips the IBM way at its existing fabs including the one in Malta that is conveniently close to IBM’s research and engineering facilities. The landscape looks, to an outsider, like one with at least two routes to IBM’s next generation of Power and z mainframe chips, if IBM got out of chip baking but continued to design its own circuits.
IBM would, if it agreed to a disposal deal by the end of the year, undoubtedly embark on a campaign to reassure users of its proprietary processors that they could expect Big Blue to retain its leadership. But these same users would be inclined to imagine, whether or not the press carried speculation on the topic, that IBM was bound to be examining strategic alternatives including the use of Intel and ARM processors to power p and z systems.
IBM’s disposal of the fabs, despite top notch work by the company’s spin doctors, would be a two-minute warning for some of IBM’s biggest customers. The mainframe crowd would probably guess that its hardware is more vulnerable to a technology shift than Power. Systems experts who know how IBM big iron really works constitute an aging, dwindling crowd, except maybe in India. Mainframes do take advantage of various processor hardware features, but are well abstracted and thus portable. The long and lively history of IBM-compatible hardware and software systems is one of the most entertaining tales in computing. Big Blue’s mastery of the technical, marketing, financial, and legal defenses required to preserve its control of the mainframe market is unmatched in information technology and arguably in all of capitalism.
But the game may be ending soon. That’s what the warning means. And whether the next season will be played on the same fields under the same rules is anyone’s guess.