Server buyers who have been waiting for IBM to ship the Power6+ chip will be surprised to discover that it already has. It just didn't tell anyone.
Remember the Power Systems revamp Big Blue did back in October 2008? Well, guess what? Several of those machines were based on the Power6+ processor, the kicker to the dual-core Power6 chips first announced in July 2007. The clock speeds didn't go up, and in some cases, clocks were turned down so that IBM could pack more processors into a single box.
The 18 to 24 month Moore's Law curve led industry watchers to expect a Power6+ kicker about right now, and as I explained last week, Big Blue's own roadmaps suggested that a kicker was coming. Roadmaps from late 2006 showed IBM expecting a 4 and 5 GHz Power6 for 2009, and this was described as a "High Freq" and "Multi-Core" kicker called Power6+ with twice the oomph of the Power6.
By early 2007, the roadmap coming out of IBM showed a 2009 Power6 coming clocking from 3.5 GHz and 5 GHz, with Power7 "to come." IBM stopped being precise right about here. Way back when, as Power5+ was first being delivered in 2004, the word on the street was that the Power6 and Power6+ generations would span from 3 GHz to 6 GHz, and given that IBM was planning on an "enhanced transistor for higher frequencies," as it said in its roadmaps, it was reasonable to expect Power6+ kickers ranging in speed from 4.5 GHz to 6 GHz.
That clearly has not happened. Scott Handy, vice president of marketing and strategy for the Power Systems division, said in an interview that in later plans than I was working from last week, IBM was actually planning to save 5 GHz clock speeds for the Power6+ chips as well as some other features. But then it discovered that the speeds of the Power6 chips could be pushed up to 5 GHz as the company transitioned to a new instruction pipeline and moved to a 65 nanometer copper/SOI process.
So instead of 3 GHz to 4 GHz for Power6 chips, some pushed up toward 5 GHz or hit it, depending on the machine.
So why not admit back in October that the Power6+ chip was out? (Some tech specs online on that October launch day said the chips were Power6+, not Power6, and I was told these were typos by IBM's PR team). For the same reason that IBM isn't going to talk about Power6+ in the server announcements it will be making Tuesday at its Dynamic Infrastructure server shindig. Everything will probably be scrubbed of pluses because IBM doesn't want to field questions about the expectations of increasing performance.
"We are continuing to take share and we didn't need to highlight performance," Handy explained in reference to the decision to not admit that some of the October machines were based on Power6+ chips. "We do think that more of the value in our Power Systems is coming from the Power software stack."
This is called the "performance plus" marketing strategy at the Power Systems division. Rather than focus on the relentless pursuit of performance, IBM wants to focus on the other attributes of the system to sell against Hewlett-Packard, Sun Microsystems, and Dell, among others.
Besides, Handy added, in the heavily virtualized environments where IBM is peddling Power-based servers, CPU speed is not as limiting a factor as is main memory capacity. And IBM is able to cram 256GB into a 4U Power 550 (which sports from 2 to 8 cores) and up to 384 GB in the same 4U space with the Power 560, which was one of the machines using a geared down Power6+ chip running at 3.6 GHz and which had from 4 to 16 cores in that box. (Each Power6 or Power6+ chip is a two-core chip with two threads per core).
What's the difference?
The Power6+ chip does have one important distinction, according to documents I have cadged from inside Big Blue. The Power6 chip had a total of eight memory keys (seven for the kernel and one for the userspace for applications), while the Power6+ chips have a total of 16 memory keys (eight for the kernel, seven for userspace, and one for the PowerVM hypervisor). According to the IBM documents, the additional eight memory keys "help prevent accidental memory overwrites that could cause critical applications to crash."
Presumably, this sort of thing is not happening on prior Power5 and Power6 chips that do not have the extra memory keys. And by talking about this as a key feature in a press release last October (if IBM had admitted the chips in the machines were Power6+ and brought up the topic), it would have been a delicate matter to bring up that the older chips didn't have them.
I can't help but believe IBM had a lot higher hopes than this for the Power6+ chip, but without Big Blue making statements publicly about its plans and without more current roadmaps, it is hard to say what happened between early 2007 and late 2008. And IBM is happy to keep us all guessing.
And Handy was not interested in talking about the possibility of a real clock crank on the Power6+ chips before the octo-core Power7 chip is delivered in early 2010 using a 45 nanometer process. Handy gave the standard "you know I can't talk about unannounced products" when I asked about this. But then again, it seems that IBM doesn't always talk about announced products, like the Power6+ chips plunked into several Power Systems machines last fall.
It seems likely that IBM will do something to boost clock speeds on the Power6+ chip in the fall and get it into every machine in the lineup. (The 64-core Power 595 is still using Power6 chips and is still running at 5 GHz). The pressure on Big Blue to do something will come through the quad-core "Tukwila" Itanium arriving from Hewlett-Packard and Intel, which was pushed out to mid-2009 as this year was getting rolling. Tukwila is a year or more behind schedule, just like Power6 was and Sun Microsystems' 16-core "Rock" UltraSparc-RK chip is at least a year behind, if not more.
If by some miracle of accounting Sun Microsystems or Oracle - if the $5.6bn acquisition of Sun proceeds as planned - get the "Supernova" servers using Rock into the field and they show good benchmarks, it would not be odd to see a Power6+ Power 595 suddenly appear running at 6 GHz. But the real pressure will likely come from below: four-socket and larger boxes using Intel's future "Nehalem EX" octo-core chips.
Ironically, these are made by IBM as well as others. And then there's the "Istanbul" six-shooters from Advanced Micro Devices that will eat into the Power 570 and low-end Power 595 space too - if IBM isn't careful. If there is even a slight delay with Power7, IBM will crank up the Power6+ clocks and hope customers can take the heat.
Anyway, we'll keep you posted Tuesday on whatever Power6+, er, Power6 machines IBM debuts as part of this Dynamic Infrastructure blitz. Don't expect to see any pluses in the press releases or spec sheets. They probably won't be there, unless Big Blue realizes that when you don't tell customers what is really happening, it looks like you are hiding something. And if you look like you are hiding something, in my experience, that usually means you are. And sooner or later, people find out what. ®