Come On Out, Power6+, You Win
April 20, 2009 Timothy Prickett Morgan
When it comes to the Power6+ processor, IBM has been as quiet as a child playing hide and seek that wants to win the game so badly that she ends up falling asleep in the hall closet where the sheets and towels are and can’t hear that everyone has given up the game. This is a great way to win the game–my wife actually did this as a child–but it also has the effect of scaring the living daylights out of the adults.
The Power6+ chip is, of course, the presumed kicker to the dual-core Power6 chip that IBM launched in the summer of 2007 in high-end boxes and rolled out across the unified Power Systems line this time last year. I have been watching Big Blue for a long time now, and the company has stopped bragging about technology way ahead of time–as it did with the Power4 processors at the turn of the millennium some 18 months ahead of the launch of those dual-core processors–and spends most of its time bragging about market share stats in the Unix space. When people know a new machine is coming, they stop buying. Moreover, the Power6 chips, which were originally scheduled for 2006 with kickers in 2007, didn’t make it to the field until July 2007 and were not across the product line until April 2008. The only thing worse than an impending product announcement in terms of its effect on sales is a delayed product that extends the impendingness of a product announcement for like six months or a year. So, IBM has decided to simply not talk about the future, just in case something goes wrong–unless you want to discuss services or some business far removed from servers, like smart water meters.
Whatever Power6+ is or isn’t remains unclear, as does its launch schedule. I am beginning to think the Power6+ is some figment of my imagination, or a hallucination, particularly with the eight-core, hybrid architecture Power7 chip due early next year. But it does exist. Here’s a 2005 roadmap that shows it:
IBM went off the road a little bit on that roadmap from 2005, and in late 2006, this is the Power Systems roadmap the company was showing customers. As far as I know, this is still accurate:
I have been arguing over the past two weeks that Big Blue needs to do something to allow the Power Systems i better compete against Intel‘s quad-core “Nehalem EP” Xeon 5500 processors, which are used in two-socket X64 servers. Now would be a good time for Power6+ to come out, with Intel’s Nehalem cards on the table, Advanced Micro Devices not expecting to put its six-core “Istanbul” Opteron chips in the field until late this year, Intel having delayed the quad-core “Tukwila” Itaniums until June or July with system shipments probably later than that, and Sun Microsystems looking weaker every day for reasons that only make sense to Wall Streeters who paid too much for Sun stock a few years back. (Sun is a good company with some great technology, and one that simply has too many employees.)
IBM likes to do summer and spring, sometimes fall, Power server announcements, but the poor global economy and the merger of the System p and System i product lines have pretty much killed anything resembling tradition in midrange server launches. I’ll need more data if I want to establish a new pattern.
Just because IBM is being cagey about Power6+ doesn’t mean things are not out there on the Web that point to it. First and foremost, since about 2005, the last page of an awful lot of IBM documents–the place where IBM covers all of its zillions of trademarks–has had Power6+ as a trademark. So IBM has been clearly been planning on launching something by that name. And soon, we’ll be seeing Power7 trademarked, too.
This link on the IBM Research site (which I can only see in Google cache because IBM has put password protection on this part of its site) that puts a name to the Power6+ chip: Eric Fluhr. According to his biography, Fluhr works for Systems and Technology Group’s Austin, Texas, lab; after getting a BS in computer science and MS in electrical engineering at the Georgia Institute of Technology 1996, he started working at IBM Austin on the Power3 chip in custom and array circuit design and circuit logic verification. Late in the Power4 chip effort, according to this bio, Fluhr switched to load/store logic design, a job he did through the development of the Power5 chip and by Power6, he led the load/store physical design team. Here’s the important bit: Fluhr is the microprocessor technical leader for the Power6+ chip. So we are not hallucinating.
In the i world, Power6+ is starting to pop up in PTFs for the i 6.1 operating system as well, as in here and there. (I don’t claim this is the exclusive listing, mind you. This is what I could find.) And on the p side of the house, here’s a reference that talks about various processor compatibility modes that are necessary to support Partition Mobility, an AIX feature that allows for a logical partition to be live migrated from one physical System p or Power System box to another one.
What no one seems to know–or not anyone who will talk to a wisecracker like me–is what the Power6+ chip will be. A radical design change seems unlikely. That’s not the way IBM does things. And it doesn’t look like IBM will do a chip manufacturing process shrink from the current 65 nanometer processes to 45 nanometers, although you would think it would want to test out the 45 nanometer processes before Power7 puts them in production. The roadmap above from 2005 suggests that Power6+ would come with “enhanced transistor for higher frequencies,” something that was removed from the late 2006 roadmap which nonetheless suggested that Power6+ would be a “high freq multi-core” chip. That suggests to me that IBM will try to crank the clocks with Power6+, maybe as high as the 6 GHz target range the company was originally shooting for with Power6 (as far as I know), but the multicore wording suggests IBM might be boosting the core count above the two used in Power6. I haven’t said this before, but it is also possible that IBM will boost the number of virtual threads in the Power6+ chip. It could, for instance, keep the cores at two per chip, but boost the threads from two per core to four.
This would not be the first time a vendor took this tack. Sun launched its “Niagara” T1 processors with eight cores and four threads in each core, then boosted that to eight cores with eight threads each with the T2 chips. The future Niagara T3 chips are expected to have 16 Sparc cores, with each core having a staggering 16 threads each, for a total of 256 threads. Pop four of these Niagara T3 chips in a box, and you have 1,024 threads–eight times as many as the top-end Power 595 has today.
Anyway, what we know is that Power6+ is supposed to have twice the performance of Power6, and four times the oomph of the original Power5 chips from 2004. Adding two threads per core and boosting the clock speed to 6 GHz should do the trick, and it won’t require a shift to 45 nanometer processes (which IBM may have been originally planning back in 2005). Last summer, when I was thinking about Power6+, I reckoned that IBM would have to double up cores to four per chip and boost the clocks to 6 GHz to double up the performance compared to Power6. But IBM could do thread boosting and clock boosting instead. IBM could put two whole dual-core Power6 chips into a single package with Power6+, reducing the clock speed to maybe 4 GHz, and boost per-socket performance that way. But I have not gotten the impression that this is where IBM wants to go. The company did this doubling with the Power5+ quad core modules, and did so because the shift from 130 nanometer to 90 nanometer processes did not yield the clock boost IBM was expecting.
What I do know is that Power7 will have eight cores (and maybe not all of them are Power-style cores) and will be “highly threaded” according to the IBM roadmaps. And rather than test out how to cram more cores onto a die with Power6+ as a dry run for Power7, the company might be adding more threads to test out how to do that best.
No matter what IBM does with Power6+, and when, what is important to Power Systems i shops is that the i 6.1 software stack be tweaked to actually take advantage of the threads or cores or clocks and yield performance increases. And in this economy, the whole shebang had better be attractively priced. Nehalem EP server makers are charging a tiny premium for roughly twice the performance. That is something that IBM has not done since the PowerPC AS/400 s were launched back in 1995. Now would be a very good time to repeat that little bit of history and encourage customers to upgrade to new gear and software.