Start Planning for Power7 Iron Now
September 21, 2009 Timothy Prickett Morgan
IBM‘s future Power7 chip may be just about done as far as the engineering is concerned, and its server designs might also be more or less completed as well. But there is plenty of time yet to tweak the boxes, and I doubt very much that the final packaging and pricing for the future Power7 machinery is anywhere close to being set. Which is a pity, really.
With the fourth quarter nearly upon us and many companies finishing up their 2009 budgets and planning for their system upgrades in 2010, now is the time to start thinking about Power7 machinery. And yet, IBM is only now talking about the basic feeds and speeds of the processor, as The Four Hundred explained last week. If this were a normal chip rollout from IBM, we’d have been talking about the Power7 chip last spring and new machines would be coming out in October, perhaps with a September announcement. Like maybe today. But with Intel and Sun Microsystems bumbling along with their high-end chips–the quad-core “Tukwila” Itanium is delayed until early next year after the umpteenth time and Sun has quietly euthanized its 16-core “Rock” UltraSparc-RK chip for its big iron, Big Blue is just not feeling the burn to get Power7 out the door. And to be honest, it wasn’t feeling much of a burning desire to do anything impressive with Power6+, and as you know from reading this newsletter, IBM didn’t.
Which leaves you scratching your head about IBM’s future Power7 processor and Power Systems plans at exactly the wrong time.
Luckily, you are on a first name basis with Sam Palmisano, and you can get a roadmap and future server briefing, and a guarantee that future machines won’t cost a lot less than current ones, so you can feel confident in pushing through that order for a Power6+ version of a Power System to sop up your 2009 budget dollars. Or, conversely, Big Sam has let you know with a wink that IBM’s pricing will be very aggressive on Power7 boxes and that you had better negotiate hard on your current Power6+ deal, or better still, that you should wait and get a Power7 box early in 2010.
Wait, you don’t know Big Sam? Oh, you’re the perfect Power Systems customer, then.
Without any real information, that leaves you, me, and a whole lot of pencils and cocktail napkins. So, do what I do. Grab a beer and let’s do a little thought experiment and figure out what we need to worry about as Power7 looms.
1. Power6+ will probably continue to be the chip of choice in Power 520 machines, at least until the Power7 chip production ramps up. The Power 520 is the machine acquired most often by Power Systems i shops, and let’s be honest, if the box were tweaked with fatter 2.5-inch drives, eight drives are plenty enough for many i shops. So is a 4.7 GHz Power6 chip with two cores turned on, rated at 9,500 CPWs and a four-core machine (with two system boards) rated at 18,300 CPWs is plenty for green screen and modernized apps.
As I think I have said since the beginning of the Pleistocene Era, when megafauna took over and we learned how to cooperate at least some of the time instead of being crazy monkeys, what Power 520 customers using the i platform need is not more oomph, but lower prices. IBM could save itself a whole lot of development effort, and mop up a whole lot of vintage AS/400 and iSeries shops, it if would just get a complete Power 520 box into the field using Power6+ chips that cost something like $5,000. IBM should use Moore’s Law to lower the price tag on Power 520s and expand the installed base. Then wait until customers grow into bigger boxes.
I am now saying that there will not be Power7-based two-socket and four-socket boxes. There will be, and IBM has said as much, as I told you last week. IBM also said that these entry Power7 chips would have one of its memory controllers cut and would only have three 4 byte links between processor cores on the die. I would guess that the two-socket and four-socket machines will have only four cores activated, and that the six-core and eight-core Power7 variants will only be used in midrange and high-end servers, where both memory controllers will be turned on and so will external SMP links. Even if the four-socket Power7s are clocked down to 2.5 GHz or so, a two-socket machine with four cores in each socket will still have lots of oomph. It is hard to say how much, since IBM is once again rejiggering the Power instruction pipeline and each core will have four threads instead of two in the Power5 and Power6 designs. My point is that an entry Power 520-class machine using Power7 chips will pack a lot of oomph, even after IBM turns off cores and gears the clocks down from the 3 GHz to 4 GHz clock speed the top-end Power7 parts are expected to run at.
Just like IBM sold iSeries boxes with S-Star chips in them in the entry line as it was rolling out the Power4 chips in 2001, IBM could bifurcate the line again with the Power7 launch. And given customer needs and application compatibility, so long as it slashes prices, no one will care.
2. Single thread performance is probably going to go down, not up, even as per socket performance quadruples. This one is a little tougher to gauge, but if IBM is dropping the target clock speed for Power7 down to 3 GHz to 4 GHz from the original 6 GHz target of the Power6+ chips, that must mean IBM is trading higher clocks to get more cores and threads on the chip.
So monolithic RPG and COBOL on the i platform and similarly monolithic C applications on AIX and Linux platforms that are sensitive to clock speeds might not see the kind of performance boost that CPW and rPerf from IBM ratings might imply. Java and PHP applications, which know how to make use of many threads, will fare better by comparison. The lower clock speeds of the Power7 chips could be washed out by the increase in threads and the huge chunk of 32 GB of eDRAM L3 memory on the Power7 chip.
3. Power Systems form factors will probably stay more or less the same with the initial Power7 launch, with a refresh down the line offering more compressed packaging. Given the amount of performance the Power7 chips will have, it would seem reasonable for IBM to get its 2U and 4U form factors for the Power5 and Power6 generations of boxes downshifted to 1U and 2U boxes. I hope this happens, but I have my doubts. IBM seems to be trying to do as little work as possible.
For larger machines, IBM has already said that it would offer upgrade paths from Power 570 and Power 595 machines using Power6 and Power6+ and preserve serial numbers, so it seems unlikely that the Power7 versions of the 570 and 595 will look much different. I also have my doubts that IBM will do blade servers much differently, so I expect the Power-based blade servers, which bear the JS designation, to have two or four sockets. The four socket blade will no doubt be a snap-together, like the Power6+ JS43 is really just two JS23s linked by a special SMP socket.
4. IBM will do everything in its power to keep base system prices the same and add in capacity to justify that stable price; that leaves the company emphasizing the increase in price/performance. We’ve seen this movie before. Like a zillion times. When an IBM marketeer cuts price, it is like admitting failure and conceding that prices have not always been perfectly reasonable. That’s why you rarely see this at IBM, or any server maker these days, for that matter. The price points stay more or less the same.
From a practical standpoint, this means you end up paying for the capacity you will get around to using many years from now today, at today’s dollar or euro. Considering the time value of money, that’s stupid. But IBM, as a creature of Wall Street, has to live quarter to quarter, and hence Big Blue, like other public server makers, has to hold those price points. So, that means something like an $11,000 entry price point for a Power 520-class machine, with $25,000 being the smallest usable configuration, to give just one example. The box might have 50 per cent more CPWs, but the clocks are going to run slower, too.
It probably means a Power 595-class machine that has maybe three to four times the aggregate performance of the current Power 595 will cost two to three times as much as that Power 595.
If you hear anything about the future Power7 machines, pipe up. Don’t be shy. And if you have any better ideas about how this might shake out, don’t hesitate to contact me. Together, we can probably figure this out.