The Little Power7 Engines That Could–And Those That Won’t
October 4, 2010 Timothy Prickett Morgan
The System/38 languished as an exotic and expensive box through most of the 1980s and the System/36 entry systems sold like hot cakes a few years later, and that experience taught IBM two lessons that were embodied in the first AS/400s from 1988 and continue to be the hallmark of Big Blue’s midrange systems today, regardless of their processor architecture. The first is that it is necessary to have as broad a selection of systems as can be affordably designed and manufactured, and the second is that you can always charge a premium for more performance and greater scalability.
In last week’s analysis of the Power 720 systems, I focused on the per-user costs of these new Power7-based entry systems and how they compared to prior 520-class Power5 and Power6 iron. In general, the cost per user came down pretty fast but then has stabilized such that it is still quite expensive to buy an entry machine for only a modest number of users; the cost per user obviously goes down quite a bit if you have modest computer needs but lots of users.
What I did not discuss in detail–and yet what was available in the detailed chart I made for the past several generations of machines–was how the cost of a unit of raw computing power has come down dramatically. So when you look at the comparisons I have made for the Power7 blade servers, the Power 750, and the now the Power 720, take heart in the fact that if you have CPW-intensive workloads–CPW being short for IBM’s Commercial Performance Workload relative benchmark rating used on OS/400 and IBM i workloads–and the number of users is less of a concern, then IBM’s successive generations of Power-based systems have given you better bang for the buck. (See the Related Stories section below for those comparative analyses.)
Knowing this, however, does not help you figure out what Power7 machine to buy today. There are myriad entry machines that overlap in performance with midrange boxes, and the midrange boxes overlap with the enterprise-class boxes, which in turn have some overlap with the big bad Power 795. The Power 780 and Power 795 have TurboCore modes, where half the cores in the box (but not the cache) can be turned off and the clock speeds can be jacked up a bit, which can help certain workloads run faster. IBM does not offer the same processors and processor feature cards across the line, and pricing on processor cards and core activations are far from uniform across the Power7-based Power Systems lineup? How is an IT manager to choose?
Very carefully, as it turns out.
To help you get a handle on your options, I have built a standard TPM monster table (patent pending) that shows all of the processor options available in the current Power7 lineup that have CPW ratings. As you know, some of the larger configurations of the Power 795 are useless unless you want to partition up the box, since IBM i 7.1 only scales up to 32 cores and 128 threads and requires a special set of PTF patches to double up to 64 cores and 256 threads. (See this story for more on that.) In this processor feature comparison table, I show the processor feature card, its base price, the per-core activation fees for the processors, and the CPW ratings for the number of cores where IBM has provided CPW test results. I then did a little math, figuring out the cost per core for each loaded up processor feature, the cost per CPW the feature card delivers, and the effective CPW per core based on the CPW results for the entire feature card. In some cases, IBM makes you buy multiple feature cards (such as in the Power 730), and in others, you can start with one and add more if you need it (such as in the Power 740).
Note: none of these comparisons include the cost of the server chassis, memory, disk, or other features or the cost of licensing the IBM i systems software. This is just a raw processor comparison. In general, as I have explained in past stories, entry machines have very affordable IBM i licenses ($2,245 per core for entry machines), but user licenses (at $250 a pop) can add up quickly to a lot of cash if you have lots of users on a small system. IBM licenses are very pricey at the high end ($53,000 per core without Software Maintenance for a year, which runs $6,000).
The first thing you will see as you dance around the monster Power7 CPW processor table is that the performance per core hovers around 6,000 CPWs, going up or down with clock speed and symmetric multiprocessing overhead as more feature cards are added to the fatter systems. But the price of that performance varies hugely depending on the machine and form factor you choose. To give you a simple visual image of how the different servers stack up in terms of their raw bang for the buck, take a look at this pretty little chart:
This chart shows the average cost per CPW for the machine; where there are multiple processor features available, I took an average. (The table has the raw data, which wiggles around a lot within different Power Systems models.) The three machines in the green are the ones that are most likely to be acquired by IBM i shops. And I do not think for an instant that the slightly elevated cost per CPW is an accident. I think it is a plan. IBM has priced the Power 710 and 730 for its AIX and Linux customers, and the PS700 blade was clearly put out to appease entry IBM i shops. But I think the blade is too expensive, as I said back in April, compared to the PS700 blade, and this chart shows I am right. The PS700 should be so cheap no one even thinks about it.
The other thing that is immediately obvious from this chart above is that it is very expensive on a per CPW basis to buy processing capacity on Power 780 and 795 boxes. These big boxes have faster clocks and lots more main memory and disk capacity, and charging more for processing capacity is a way to compensate for the fact that entry systems have the same disk and memory prices and larger boxes have higher but also uniform memory and disk prices. IBM could charge the same for processors on a per-CPW basis and charge varying amounts for main memory and disk, but that would only encourage clone memory and disk makers to swoop into the market as they have in the past. The odds of someone making a clone Power7 chip are precisely zero. There are more odds that I will sprout wings right here right now, along with my wife who like myself is working late, and we will be able to fly away to a lovely vacation on a pretty green planet circling Alpha Centauri. So IBM can get away with charging 15 cents or so for a CPW on a PS700 or PS702 blade but over $2 on a Power 780 or 795 enterprise-class servers.
It is important to drill down into the data, of course, and not make too many generalizations. Take a gander at this chart:
Now, in this case, I have taken the most likely IBM boxes–the PS700 blade and the Power 720 and 750 rack/tower servers–and zoomed down into their different processor features. All processor-feature card combinations in the Power 720s and Power 750s are not created equal, obviously. But look at how linear the price increase is. And look at how the fastest, most scalable card in either model always goes a little north of linear. This is a pattern we have seen since the AS/400.
The funny thing about the pricing is that it is probably a fairly good predictor of where IBM is getting good yields on its processors and where it is not. In general, the faster the processor clock speed or the more cores on a chip that are activate-able, the more expensive the processor feature. The eight-core, 3 GHz and 3.55 GHz Power7 chips used in some of the entry Power Systems machines are just not worth the money unless you need those cores and clocks.
If your applications are multithreaded, you can use lower speed chips with more cores and get the same work done (in theory) and save yourself some money. As far as I know, the DB2 for i database already knows how to make use of the threads in the Power7 machines, so you don’t have to do anything extra special. (Remember, the Power7 chips have four threads per core.) The real question you need to ask your programmers and IBM is can you get by with a 3 GHz processor today, thanks to all that extra on-chip cache and larger main memory, when you are moving from older Power6 and Power6+ machines spinning at 3.7 GHz to 5 GHz. If you are moving from Power5 and Power5+ chips, you have similar issues, but it is hard to believe that the Power7 cores at nearly any speed–even the slowest 3 GHz–won’t stomp these older chips into dust.
I hope you find this useful, and happy shopping. I will be putting together Unix and Windows comparisons for the entry Power7 machines in the coming weeks to see how they stack up competitively.