IBM’s Evolving Power Systems Rollout
July 12, 2010 Timothy Prickett Morgan
Shooting and hitting a moving target is never easy, but nailing a moving target when you are flying down a bumpy road trying to hit that target is another thing entirely. Which is why the server business is a bit like the chase scene in a thriller movie. The targets that vendors are aiming at are chip and system roadmaps, and the blood is red ink, but there is a certain kind of drama in it. Particularly if your company’s business, and perhaps your job, depends on the reliability, scalability, and availability of your primary server platforms.
I do plenty of complaining about and cajoling of all of the server makers, not just IBM. But The Four Hundred is predominantly dedicated to the AS/400 platform and its progeny, which means in this newsletter at least I spend most of my time analyzing what IBM might do, suggesting what might be done, analyzing what has been done, and then closing the loop by suggesting what might be done even better. We’re in the middle of the Power7 loop at the moment, awaiting low-end and high-end machines, and the Power7 plan has, as far as I can tell, changed a number of times and pretty rapidly as economic conditions and IBM’s own Systems and Technology Business dictated.
The change in IBM’s plan is not necessarily a bad thing. In fact, I think IBM is doing a better job with the Power7 systems launches of reacting to market conditions. And for selfish reasons, of course, because it is Big Blue’s job to maximize revenues and profits. To see how things are changing, however, you need to see where they were supposed to end up originally. Depending on who I asked late last year inside of IBM, I heard initial Power7 boxes were due in January or February, perhaps in March if yields were good; others said it would be May. Things were in flux at the end of December and in early January because IBM had a better sense of what Power7 chip yields would be, had put patched versions of i 6.1, AIX 5.3, and AIX 6.1 into the field that could support Power7 processors, and the economy looked like it might be on the mend just at the time when it was pretty clear that Power Systems shops would curtail spending until Power7 boxes came out in 2010. I don’t know when this server roadmap illustrated below comes from precisely, but I would guess it was early January. Take a look:
You may not know the IBM lingo for Power-based systems, so let me translate. In the April 2010 column, the machine designated as the P7 MR16/64 is none other than the Power 770, which scales from two to eight sockets in one to four chassis; the Power 780 is just the same box with TurboCore mode activated. MR is short for Mid Range, and it is the designation that IBM has been using for years for X70-class machines. The P7 HV32 is the four-socket, 32-core Power 750 machine as well as its supercomputing variant, the Power 755, which comes with only 32 cores activated. (HV is short for High Volume.) The plan back then was clearly, according to this roadmap, to debut these machines along with a new Technology Level patch for the AIX operating systems in April, most likely with shipments in May.
As you know, that didn’t happen. The Power 750, 770, and 780 machines were pulled forward into early February, ahead of the onslaught of the “Tukwila” Itanium 9300, “Westmere-EP” Xeon 5600, and “Nehalem-EX” Xeon 7500 processors from Intel and the “Magny-Cours” Opteron 6100 and “Lisbon” Opteron 4100 processors from Advanced Micro Devices.
According to this roadmap above, IBM had four other classes of Power7 machines in the works for 2010, all due in October. In Power lingo, HE is short for high-end and has been used to describe the Power 590 and 595 machines. The IH node has always been used to describe the 575-class, special purpose supercomputer nodes, and in this case, this is none other than the IH node that is to be used in the petaflops-class “Blue Waters” massively parallel supercomputer going into the University of Illinois later this year. (I told you about it last fall here, when I saw it at the SC09 supercomputing conference.) The Power7 blades were also due in October, but were announced in April as the Power Systems 700, 701, and 702 blades, as you well know.
There was another machine called the P7 HV16, which means a High Volume Power7-based machine with a maximum of 16 cores activated. This is what I have been tentatively calling the Power 720, the expected kicker to the Power 520.
That was it for the Power7 systems rollout as the new year got under way.
Like many of you, I have been lobbying IBM for some time to do more than just put a Power 720 into the field, mainly because if a four-core Power 520 was overkill for a lot of i shops, then a Power 720 with eight or 16 cores is going to be absolutely silly, given that the performance per core will be better, not worse, even with the lower clock speeds compared to Power6 and Power6+ cores. (I had not foreseen that addition of 32 MB of eDRAM to the Power7 chips, and hence had been warning that if clock speeds dropped, performance on many workloads would. I never claimed to be perfect, and I am usually working from very thin data. But that on-chip cache made up and then some for the lower clock speeds on the Power7 chips.)
As we learned from Ross Mauri, general manager of the Power Systems division within IBM’s Systems and Technology Group at the COMMON midrange user group meeting in early May, the Power7 server roadmap is a bit deeper than what the roadmap above shows. Mauri didn’t give out a lot of details, of course, but he did say that IBM has not one, but four different entry Power7-based systems in the works.
If I had to guess, I would say there is a P7 HV4 and P7 HV8 box in the works in addition to the P7 HV16, with one of these tweaked to make a Smart Cube appliance. For all I know, there is only the P7 HV16 box and the other three are geared down Smart Cubes. (I hope not. We’ll see.) The HV16 machine, which I have been calling the Power 720 as IBM probably will, should have 16 Power7 cores, which means it is a two-socket box and is essentially half of a Power 750. Cut the Power 720 in half, you get a Power 710 with eight cores, and maybe this is a single-socket machine, like the PS701 blade, and maybe it is a two-socket box using half-dud Power7 chips that would be tossed out. (I would do the latter.) It would be interesting to cut this machine in half again, putting in just four cores and one socket, making what IBM might call the P7 HV4 internally and what I would call a Power 705. This Power 705 would probably make a good kicker to the current Power 520-based Smart Cube appliance.
The reason I bring any of this up is to show that IBM isn’t just chiseling a roadmap in stone and then sticking with the plan, no matter what. There’s been plenty of wiggle, and I think that is because Mauri and his team are getting lots of feedback from the i community, and they are listening. The AIX folks generally do not give a care about entry Power systems. That market was ceded to Windows and Linux years ago. That said, it would be very interesting to see IBM put cheap AIX clusters on the market, perhaps using revved up clock speeds on half-dud or even three-quarters-dud Power7 chips. Imagine if you had a chip with two Power7 cores running at 5 GHz or 6 GHz, maybe with 8 MB or 16 MB of eDRAM cache, and then plunked two of them into a cheap $2,000 server node aimed at HPC customers. . . .
Whatever IBM is up to with the low-end and the high-end, it needs to get these boxes out in the field as soon as possible, and I think wherever possible, IBM will be moving up machine launches yet again. The Power 795 launch cannot be moved up until deliveries of 256 GB memory cards are moved up, since the current 32 GB cards are not fat enough to yield a balanced system with 256 cores and 1,024 threads. This is why IBM has only tested Power 770 and 780 machines with one quarter of their cores activated on most benchmarks; it needs the fat memory cards, which were not expected until November. Moreover, AIX 7.1, which will be needed to scale across 256 cores and 1,024 threads, is not going to be ready until October. You can’t hurry software like you can hardware. If AIX 7.1 is ready and the fat memory cards are ready, then the Power 795 could move ahead, too. Maybe even in July or August.
That leaves the original Power 720 machine, a.k.a. P7 HV16 on the roadmap, and whatever other three machines Big Blue has cooked up since early this year after listening to complaints from customers, partners, and curmudgeons in the Power Systems i community. And I think if it is physically possible, these machines–whatever they are–will come out well before October. IBM wants to have boxes in the field to sell against Hewlett-Packard‘s HP-UX and Oracle‘s Solaris platforms. HP is ramping up its Integrity blades using the Itanium 9300s–the BL860c i2 (two-socket, single-wide) and its double-wide and quadruple-wide SMP siblings, the BL870c i2 and BL890c i2–and is readying a rack server using the chips called the rx2800 i2, with two-sockets. Oracle is supposed to be getting its “Rainbow Falls” 16-core, 128-thread Sparc T3 processor into the field in the coming months for machines with one, two, or four sockets; this chip will spin at around 1.67 GHz, so while it has lots of cores and threads, they don’t move very fast. Still, IBM needs entry machines to compete against this Itanium and Sparc iron and to appease i shops with modest computing needs.