Midrange Power7+ Servers: Scaling Up Means Bring Your Checkbook
March 25, 2013 Timothy Prickett Morgan
We have spent several weeks walking you through the salient technical and economic characteristics of the new entry Power Systems machines using the new Power7+ processors from IBM. This week, it is time to take a run at those midrange Power 750+ and Power 760+ boxes that are based on the double-stuffed dual-chip modules, or DCMs, that are used to put more cores and threads in these boxes than Big Blue might otherwise not be able to deliver using regular single-chip module, or SCM, units that are used on the entry machines.
If you want to review the feeds and speeds and slots and watts of the Power 750+ and Power 760+ machines that were announced along with the midrange boxes back on February 5, check out this story. I’ll give you a brief recap.
The chassis for this box, which is only available as a rack server, is 5U high and is a completely new design that is aired out to accommodate more cooling for processor cards and memory sticks. Both machines have up to four processor sockets, with the Power 750+ using DCMs that have eight cores (that’s two four-core chips sharing one socket) and the Power 760+ using DCMs that have 12 cores (that’s two six core chips snuggled into a socket). The Power 750+ tops out at 1 TB of main memory, while the Power 760+ tops out at 2 TB, which is around what other four-socket servers can do. (The Xeon E5-4600 processor from Intel can do 1.5 TB across two sockets.) Aside from cramming more cores and memory into the same chassis and having substantially higher prices, the Power 760+ has one other thing that the Power 750+ does not, which helps to justify its cost somewhat. And that is the ability to do Capacity Upgrade on Demand (CUoD) processor activations.
Basically, those entry Power7+ machines–the Power 710+, 720+, 730+, and 740+ servers–are being pitched against servers using Intel’s Xeon E5-2400 and E5-2600 processors, which are used in two-socket servers, and the midrange Power7+ boxes–these Power 750+ and 760+ boxes–are lined up against the four-socket servers using Xeon E5-4600s. Obviously, these respective Power7+ machine families are also designed to compete against alternative systems based on Advanced Micro Devices‘ Opteron 4300 and 6300 processors and the comparable Sparc and Itanium systems from Oracle, Fujitsu, and Hewlett-Packard.
Here are the memory capacity and processor options for the Power 750+ and Power 760+ machines, laid out in an easily readable format. (Well, that was the design goal, anyway).
In the table above, I show the different processor options and the maximum number of cores at each speed. The price is the cost of getting that particular processor feature card and activating the number of cores on each card as shown across the top–8, 12, 16, 24, 32, 36, or 48. The performance is reckoned in Commercial Performance Workload (CPW) units, as has been the case for OS/400 and IBM i workloads since Big Blue retired the RAMP-C ratings a zillion years ago. The cost per CPW is just for the processor feature card, not for any other part of the system. I wanted to show the inherent cost of the processor card and activated cores only at first.
As you can see, the lower-clocked Power7+ chips are considerably less expensive than the higher-speed chips used in either machine. With IBM i shops, because of the fact that the operating system and its integrated database is priced per core, you want to try to get the most performance per core as you can. But, but, but, as you can see, the raw compute capacity in the 4.0 GHz DCM used in the Power 750+ costs about twice as much per CPW as the 3.5 GHz option, and ditto for the 3.4 GHz DCM option in the Power 760+ compared to the slower 3.1 GHz chip modules. In other words, it is very tough to win.
As you scale up the processing within a single machine and processor option, the cost per CPW rises a bit, but that is entirely due to the overhead of SMP scaling on a machine. Four times the processor count yields between 3.53 and 3.7 times the oomph on online transaction processing workloads. IBM has come close to not charging for that missing performance that gets used up by SMP overhead, but it didn’t quite get rid of it all. (How intentional this all is, I cannot say. It may be a coincidence. But I suspect not, given the number of significant digits in the processor card and core activation prices for these machines.)
So how do configured machines stack up? Good question. And here’s a good answer in this handy table right here:
Just to give you a frame of reference, I rolled the configurations and pricing of the Power 740+ machine into this table alongside the Power 750+ and Power 760+ machines. As you can see, in many cases, the midrange boxes yield about the same bang for the buck configured up with processors and a reasonable amount of main memory. These configurations also include a RAID 5 disk controller, enough 856 GB 10K RPM disks to give it a reasonable amount of base capacity, a storage backplane for those disks, a quad-port Gigabit Ethernet adapter, and a power supply. There are other things you need to add to the machines to make them useful in many cases. But here, I want to get a sense of the base system hardware costs.
The Power 740+ and Power 750+ are in the same P20 software tier, so comparisons there are fair. But the Power 760+ is a P30 box, so you pay a big jump in IBM i and other software costs that are based on IBM’s software tier if you pick the Power 760+ over the Power 750+ machine. IBM i is $40,000 per core on a P20 box and $53,000 per core on a P30 box. (I did not put the operating system and Software Maintenance costs in these configurations. We’ll get to that in next week’s issue.) That’s a 32.5 percent increase in software cost per core jumping up that one tier.
But that is not the only incremental cost you will see. The base 760+ unit costs 14 percent more than the Power 750+ base unit (they are virtually the same, I think), and for reasons I cannot fathom, the exact same memory, disk, RAID adapter, backplane, net adapter, and power supply features I used in the configurations all cost precisely 24 percent more with the Power 760+ than they do on the Power 750+ system. That is your CUoD upgrade and memory scalability tax right there, people. Which is fine, IBM should be compensated for that capability, but I am just pointing it out. (I hope that disk, memory, and other features can transfer from the Power 750+ to the Power 760+ if you do an upgrade without a 24 percent tax.) I am not sure if every feature is more expensive on the Power 760+ than on the Power 750+ machine–there is not enough time in the week to figure that out. But I will tell you to be careful if you are trying to decide between the Power 750+ and the Power 760+ machine.
The thing is, of course, is that the Power 760+ costs more, but it also scales further, topping out at 274,000 CPWs instead of 208,000 CPWs like the Power 750+ does. And I haven’t done all of the math yet, but based on history, I am pretty sure the Power 760+ will be considerably less expensive than the Power 770+ when it comes to hardware. They have the same per-core software price, of course, but the Power 770+ cores clock higher. Here we go again. . . . It is very tough to win.