That Perplexing Power7+ Processor
October 17, 2011 Timothy Prickett Morgan
Like you, I only jump to the wrong conclusion when I come to a cliff and generally only if I believe I am going to make it to the other side. There was a lot of circumstantial and graphical evidence that IBM was going to launch a Power7+ kicker to the current Power7 processor, and as we all found out last week, that didn’t happen. As we report elsewhere in this issue, IBM put out new systems with denser memory and new PCI-Express I/O peripheral slots, but with the same or essentially the same Power7 processor features.
I hear a lot of things in my line of work, and I have learned to take it all with a grain of salt. But I heard from some sources inside IBM that there were Power Systems announcements coming in October or November, and then caught wind of the updated Power7+ and Power8 roadmaps, and added 2 and 2 and 3 together to get 7-plus. This chart, which IBM’s competitive analysis team put together after midrange system rival Oracle divulged the tech specs of its then-impending eight-core Sparc T4 processor back at the end of August, had me pretty convinced that the Power Systems machines coming out this fall would sport Power7+ chips. That, plus the fact that I can add 18 months to April 2010, when the Power7 chips were originally slated to debut for shipments in May. Anyway here’s the chart:
Best I can figure, that chart was made sometime in early September. It’s a pity I didn’t see this chart before I went off writing about planning for new systems:
You forgot my glasses and my goatee, Big Blue. And my chin is stronger than that, too.
I spoke with some of the top brass within IBM’s Power Systems division, and Steve Sibley, director of product management for the Power Systems line, conceded that IBM itself was responsible for causing the confusion. And Ian Jarman, who manages the Power Systems Software operations at IBM and who has been a product manager for Power Systems before that and for the iSeries even before that and is perhaps the single constant in the AS/400 universe within IBM for the past decade, wanted to be real clear.
“This is not and has never been intended to be the Power7+ launch,” Jarman told me emphatically.
I, of course, countered that if you do the math on IBM’s 36-month cycle for Power processor introductions and bisect it, you would expect the Power7+ chip around now. You could make an argument that it should have been here in September for October shipments, about the same time that Intel and Advanced Micro Devices were hoping to get their new Xeon and Opteron processors in the field and when Oracle did get its new eight-core Sparc T4s into the field. Oddly enough, the Xeon E5s and Opteron 6200s are shipping and have been doing so for more than a month, but neither has been launched as in formally announced.
While I welcome the doubling up of memory capacity and the move to PCI-Express 2.0 peripheral slots, there was no reason why IBM could not have supported 16 GB memory sticks in the new machines, and given that other server makers were supporting PCI-Express 2.0 peripheral slots two or three years ago, I had to do a double-take to realize IBM had not already done this on the Power Systems line. Regardless of what the official, most recent plan is, my take is that Power7+ chips should be here now so Big Blue can blunt the attack on its Power Systems business, particularly when others are holding back and particularly when IBM i shops benefit most from higher clock speeds thanks to the way IBM prices the operating system/database combo.
Neither Sibley nor Jarman wanted to talk about what the Power7 launch plan is, and that is their job to not say anything too specific. What Sibley said in May was that a Power7+ chip was in development and would be delivered in the next 12 to 18 months. Based on the chatter and charts I saw, I assumed that IBM moved up the Power7+ launch to October, and I will say further that I think the four new entry Power Systems machines were actually designed to hold Power7+ chips as well as Power7 chips since the two will no doubt be socket compatible. I think we’ll see the Power7+ chips in the line next year because IBM has to do something before the Power8 chips ship in 2013 if the roadmaps are correct. (They may no longer be, so don’t bet your budget on it.)
What IBM is telling business partners is that the old Power7 machines and the new Power7 machines are essentially the same except for the memory and I/O capacity differences–including essentially the same price. They have the same software editions riding on top of them and for most customers, according to Big Blue, either machine will work fine. Those who have high-bandwidth networking and storage needs will want the newer machines, or those that are doing lots of virtualization or other memory-chewing workloads. At around $200 per GB, that extra memory is not cheap, so not everyone will want to go there anyway. In 2012, IBM told business partners, it will start pushing the fatter Power7 machines and later in the year it will withdraw the older boxes. The key thing, IBM told resellers was DO NOT DISRUPT 4Q11 SALES.
As I point out elsewhere in this issue, if you are buying one of the skinnier Power Systems from last year’s catalog, you should demand some kind of compensation. There’s no way the older machines are of the same value on the street with smaller potential memory and slower I/O. It probably isn’t much of a difference, but there’s no way it can be zero.
More importantly, if you have I/O and memory sensitive workloads, then check out the feeds and speeds of the upcoming Xeon E5 processors for two-socket servers, which I divulged in May over at The Register. Intel is way out in front with integrating PCI-Express 3.0 peripherals into systems, which should deliver about twice the peripheral bandwidth as the PCI-Express 2.0 slots IBM just added to the new Power Systems machines. The affiliated Xeon E5 server platforms will have eight 6 Gb/sec SAS and SATA ports on the system board, and Gigabit Ethernet and 10 Gigabit Ethernet LAN ports on the system board as well. RAID 5 or RAID 10 protection is built into the “Patsburg” C600 chipset, and each Xeon E5 processor has one or two PCI-Express controllers, yielding either three or five PCI-Express x8 slots coming out of each socket and feeding through the chipsets to deliver 40 GB/sec of I/O bandwidth into each socket.
To give you an idea of what that means, this two-socket server will have as much I/O bandwidth one of the nodes in Power 770 server that has just been upgraded to PCI-Express 2.0. IBM prides itself on building high-bandwidth boxes that are worth the premium it commands, so it needs to pick up the I/O pace. Ditto for main memory, where IBM used to command a lead. These Xeon E5 machines will, in the top-end configurations, have four memory channels per socket, with three memory modules per channel. If you do the math on that and assume 16 GB DDR3 memory, you get 384 GB of main memory on a two-socket server. And did I mention that this memory will run at 1.6 GHz instead of the 1.07 GHz that IBM is using in the new and old Power Systems based on Power7 chips?
This newsletter advocates for Power Systems customers and it wants IBM to do more than worry about fourth quarter sales. It wants IBM to start taking a more aggressive technical fight to Intel so all of us in the IBM i ecosystem can do better.