Without Further Ado: Power10 Entry Server Pricing
September 26, 2022 Timothy Prickett Morgan
We have been through the new Power10 entry servers from just about every angle imaginable, which has been our great pleasure. And now, we come to the moment we have all been waiting for: What is the bang for the buck of a Power10 machine? As is usually the case when you ask any question, the answer is: It depends.
To try to figure this out, we put together pricing information for all of the different variations of processors for the Power8, Power9, and now Power10 entry servers with one of two processors. And when we say all of the variations, we mean all of them because we want to be thorough and useful.
In each case of the bare configuration machines shown, we have the Power processor feature shown with one or two sockets and a specific number of cores, all of which are activated on the machine. The Power8 and Power9 machines have four 600 GB hard disk drives for base storage and the Power10 machine has four 600 GB NVM-Express flash drives for base storage. Each machine is configured with 32 GB of memory per core, which we think of as a base amount of main memory required to run modern workloads. You can argue here or there, but we can tell you this: More memory than this can get very pricey indeed because memory is the most expensive part of any system these days – it represents about half the cost of a machine installed by a hyperscaler and cloud builder – and it looks like it will get worse instead of better because CPU compute is scaling faster than memory capacity and certainly a lot more than memory speed.
In the tables shown below, we have used IBM’s own Commercial Performance Workload (CPW) benchmark test, which is a variant of the TPC-C transaction processing benchmark, to gauge the relative performance across processor features within a Power server generation and across the three generations. In the cases where IBM has not provided CPW ratings for the processor configuration shown, we have estimated the CPW performance, denoted in bold, red italics.
For pricing, we have gotten our hands on IBM’s public price list, which is tricky to get ahold of – it’s amazing how certain links on the IBM announcement system don’t work when they have to do with pricing – but we have them. If you need them, send us an email.
IBM has tweaked the way it prices the machine in the past few generations, shifting money from the base chassis to the CPU features and core activation charges and then back to the base chassis again. We have normalized all of this, and shown the differences in the tables. The system price does not include everything you might need in your specific shop to run a workload – this is meant to be an indicator of the base compute cost. This set of comparisons does not include a single bit of software – we will take a look at that next week. This week, we are just talking about the core compute and chassis cost and the base memory and disk to support that compute.
This time around, because the cores are getting more and more powerful with the generations, we have also included a new metric for comparison: The fraction of a core that it takes to drive 30,000 CPWs of performance. This is an important factor to think about as you weigh per-core software charges within a generation of machines and across the generations.
Let’s start with the four entry Power10 machines, the Power S1014, the Power S1022s, the Power S1022, and the Power S1024:
And now, here are the three main entry Power9 machines, the Power S914, the Power S922, and the Power S924 from 2018:
We did not put in the H models aimed primarily at SAP HANA or the L models aimed primarily at Linux. These are not machines that are widely adopted by IBM i shops.
And finally, here are the three main entry Power8 machines, the Power S814, the Power S822, and the Power S824 from 2014:
The first thing that jumps out at us as we review these configurations is that the performance per core bump that Big Blue is delivering with the Power10 is huge, somewhere between 2.X and 3X that of a Power8 core and somewhere between 1.5X and 2X of a Power9 core.
The other thing that jumps out is that IBM is putting more cores inside of the box, too. The Power S1014 only has one socket with four or eight cores activated, so this is an exception to this rule. But the Power S1022s has from 4 to 16 cores, the Power S1022 has from 12 to 40 cores, and the Power S1024 has from 12 to 48 cores. These are twice as fat, in terms of core counts, as the similar Power9-based Power S922 and Power S924. The Power S914 has the option of a 6-core configuration, which is not available with the Power S1014. Ditto for the Power8 entry machines, which do not have a 4-core option as standard. (There are special versions for IBM i with fewer cores, but we are not comparing those as part of this specific analysis.)
The third thing you will note is that the price/performance improved quite a bit from the Power8 to the Power9 machines, the jump from Power9 to Power10 is, well, a little harder to qualify. Sometimes it is the same, sometimes it is less.
You will have to sort out what paths you might take from all of this. We can’t guess. But we can offer you some advice. First, within any product family – meaning Power8, Power9, or Power10 – always go for the lowest core count possible to attain any given performance level. This will reap rewards in terms of software pricing and will also boost the amount of memory and I/O available per core. Second, always go for the highest clock speed possible for any given number of cores. This will let you get the work done faster. And if you have to pick between the two, sacrifice a little performance by having slightly fewer cores that run at a slightly faster clock speed. Here is an example of what we mean here. A two-socket Power S1022 using the 2.75 GHz Power10 chips with a total of 32 cores is rated at 725,500 CPWs, and the same machine with two 20-core Power10 chips running at 2.45 GHz is rated at 789,600 CPWs. The bang for the buck is better on the faster chip and it is reasonably close in performance.
Up next, we will add in the IBM i software costs and see how this plays out.