Some Things To Ponder On The Impending Power7+ Era
October 2, 2012 Timothy Prickett Morgan
Well, the Power7+ era of IBM‘s Power Systems family of servers and now its Pure Systems modular systems is nearly upon us. Or at least the beginning of the era is. It has been a long time since IBM refreshed an entire lineup of Power-based commercial systems from top to bottom, all in one fell swoop, and there is every reason to believe that IBM will have a staggered approach to the Power7+ rollout this time around, too.
In fact, I wracked my brains, and I cannot think of when IBM last did such a rollout. Certainly not since the Power4 was put into systems back in October 2001. With the PowerPC Northstar-based AS/400 machines that came out in 1999 and the PowerPC S-Star iSeries machines that came out in early 2000, IBM offered a single processor type with a few different clock speeds and cache configurations across the entire product line. But generally speaking, IBM has always used a mixed of processor generations–or types within a generation–across its midrange and now high-end systems in the AS/400 family since 1988. Ditto for the RS/6000 line, for the most part, since it was launched in 1990. Doing a big bang upgrade across the line, where the entire line is refreshed at once and with a shiny new processor is the exception, not the rule, and even when IBM did launch entry, midrange, and high-end machines all at the same time, it used to mix old and new processors in new skins.
Rather than doing an entire product line launch once a year for each formerly distinct family of machines, IBM has instead moved systems with a lot of commonality and tries to update those machines once a quarter. This year, we got the Flex Systems in the spring alongside updated Xeon E5-based machines around the same time, with the late summer being new System zEnterprise EC12 mainframes (announced about six weeks early) and the fall being at least some portion of the Power7+ product line. Based on the lineup of speakers at the October 3 launch, which includes people on the mainframe side of the house, I would guess that this time around IBM might put the Power7+ into the Power 795 machines, and maybe into the Power 770 and 780 boxes, rather than starting in the midrange and moving out to blades, entry boxes, and then the high end as it did with the Power7 launch in 2010.
If IBM had stuck to its previous cadence for Power chip launches, and even giving it the benefit of the doubt that the plan back in early 2009 was to launch the Power7 in May 2010, not earlier in February 2010, then the Power7+ should have debuted a year and a half later, which would have put it at October 2011. And if you start counting from the initial February launch, then you would be looking for Power7+ machines to come out in August 2011. The latter would have been a silly time to do a launch, with Europe on vacation. But the fact remains that IBM is not sticking to its own cadence and I think it probably has to do with yield issues on the ramp of the 32 nanometer processes developed at its East Fishkill, New York, foundry and the fact that Intel with Itanium chips (used to run big HP-UX iron) and Fujitsu with Sparc64 chips (used to run big Solaris iron) are not exactly moving at a blinding speed. The OS/400-IBM i business plugs along and the AIX business is eating market share, and with Big Blue losing out to Cray on some very large supercomputer deals that would have put money into the Power Systems coffers, IBM is in no hurry whatsoever to get a new Power7+ chip out the door. I wonder if IBM could have gotten it out the door last fall if it wanted to, and I suspect not.
That’s unfortunate for customers, mainly because a new chip generally comes with new technologies wrapped around it and much better bang for the buck.
There are a few things to consider here. First, a processor that is late is the rule, not the exception. Intel’s Xeon E5 processors were six to nine months late. AMD’s Opteron 6200s were late. Second, IBM does the most volume in terms of server chips in the entry and midrange for Power Systems, but it makes the bulk of its revenues on big boxes. So if IBM has good yields on a relatively small number of chips with the 45 nanometer processes, the company will be tempted to start at the high-end and get a Power7+ bump in the Power 795 and maybe even the Power 770 and 780 boxes (or whatever their kickers will be called if IBM changes the names) that coincides with the expected jump in mainframe sales in the fourth quarter of this year and in the early parts of next year. Having got that wave under way, then Big Blue could refresh its blades, entry, and midrange systems, as well as the Pure Systems machines, with Power7+ boxes later.
Having said that, IBM could do something radical and only offer the Power7+ processors in revamped Flex System machines, thereby encouraging customers who need the extra oomph to move to the new modular, converged architecture. That would be truly radical, and probably not the kind of risky move IBM would make. But, if it believes that Pure Systems is the way to go, this would be one sure way to show it. (I think the odds are pretty low on this happening.)
What I can tell you for sure is that IBM will do the thing that is least disruptive to its sales for the fourth quarter and that makes the best use of whatever yields it has for Power7+ chips. And considering that there are only a few thousand high-end Power Systems shops, and IBM is on a first name basis with their CIOs, telling them the Power7+ machines are coming out mainly for them in 2012 would be an easy thing to manage and keep secret.
What has not been made clear is whether the Power7+ processors will plug into existing Power7-based machines even though they are plug-compatible with the sockets. It is also not clear if IBM will be able to offer upgrades from the Power7 to Power7+ machines. If a lot of the surrounding electronics has to change–as might be the case if IBM moves to PCI-Express 3.0 peripherals–then according to accounting rules, IBM can’t call the swap of components an upgrade and thereby preserve the serial number and existing depreciation schedule for any particular box. This being a Power7+ launch, the idea of the plus generations in any server line is to make use of the same sockets, chipsets, and systems and just use a process shrink to add more transistors and therefore features to the processor while also cleaning up the microarchitecture. This certainly happened with the Power4+, Power5+, and Power6+ chips. And there is no reason to believe the Power7+ will be any different. That said, the versions of the machine with double-stuffed sockets may require a new system design.
IBM has been mum on all this, and when it sees customers getting nervous about upgrade paths, it usually puts out a statement of direction to calm them down. This happened back in May 2006 with Power5 and Power5+ systems a year ahead of the Power6 launch. IBM wanted customers to keep buying and not wait another year for the much better Power6 chips. We have not seen any such statement. That could mean customers don’t care, it could mean upgrade paths are available, it could mean a lot of things.
There is another interesting possibility for the initial Power7+ launch: IBM could focus at the low-end and kick out machines with double-stuffed sockets in the PowerLinux line to really try to give its X86 competitors some trouble for the first time in a long time. IBM cannot use PowerLinux boxes to take a direct assault on the X86 base running Windows (roughly half of server revenues these days), but it sure can target the 20 percent of the server revenue base that loves Linux and cares less about the underlying platform than about price/performance and the availability of Red Hat or SUSE Linux distributions.
Having said all of that, if sales of Power 770, 780, and 795 machines are humming along and Xeon and Opteron machines are encroaching on the midrange and if IBM is worried about Oracle, Fujitsu, and Hewlett-Packard finding their legs in the Unix system market, then Big Blue could start the Power7+ launch in the belly of the market with the Power 750 and a Power 760 with double-stuffed sockets. This will require better yields on Power7+ manufacturing to do, of course, because the last thing IBM wants to do is launch machines and not be able to ship them.
What I can tell you for sure is this: Whatever IBM does will tell us indirectly about the yields for Power7+ and where it thinks the weakest part of its Power Systems product line is and where competitors are trying to exploit.