IBM’s Competitive Analysis For Power10 Entry Machines

August 22, 2022 Timothy Prickett Morgan

Long gone are the days when IBM, and indeed most other server vendors, clamored to run TPC-C, SAP SD, and SPEC JVM transaction processing benchmark tests on their new machines to prove they had the performance and they had the bang for the buck to compete in the broader market.

To be fair, many of those tests had their flaws, allowing a certain amount of gaming of the benchmarks (as happened with the TPC-C and SPEC JVM tests) or not having official pricing attached to them (as happened with the SPEC JVM and SAP SD tests). Still, those benchmarks, as well as a number of others, gave us a kind of baseline to begin the process of trying to compare various enterprise systems to each other. Given enough tests and some pricing data, we could figure out pricing for representative machines and also figure out their relative performance, thereby giving people like you a fighting chance to understand how a Power Systems machine stood against other RISC, Itanium, and X86 servers.

Here in 2022, every established RISC or Itanium architecture machine running either a flavor of Unix or a proprietary operating system is dead. And such systems have been dead for a decade. Yes, there are Arm RISC servers being build by Amazon Web Services and Ampere Computer, but AWS is using them for its own cloud and Ampere Computing is also focusing on selling its processors to hyperscalers and cloud builders for their private consumption and sometimes for IaaS rental on their clouds. AMD’s three generations of Epyc processors, launched in 2017 through 2021, have done progressively well in taking share away from Intel among hyperscalers, cloud builders, and HPC centers and are now making headway into OEM machines bought by regular enterprises like yours. But with a maximum of two sockets, the Epyc line has ceilings, even with 64 cores in a socket, and that makes midrange and enterprise system customers wary of adopting Epyc chips for their mission-critical, back-end systems. That leaves platforms running Intel’s Xeon SP processors as the main competition to the Power10 entry and midrange machines that Big Blue launched back on July 12.

We have done a deep dive on each of the Power10 entry machines, all of which can run the IBM i operating system, in succession, as well as talking about the architecture and performance of the midrange Power E1050, on which Big Blue does not support IBM i but we think it should. The performance metrics we have for these machines are the Commercial Performance Workload (CPW) benchmark for IBM i and Relative Performance (rPerf) benchmark for AIX and Linux, both of which are variants of the TPC-C test that Big Blue tweaked to gauge the relative performance of processor features, memory subsystems, and I/O subsystems on decades of its Power-based server machines. Data to compare the performance and price/performance of the new Power S1014, Power S1022s, Power S1022, and Power S1024 mainstream entry servers and their Linux-only Power L1022 and Power L1024 machines to servers based on the Xeon SP processor are a bit thin, but we have wrangled a presentation that Big Blue put together for its business partners and resellers that has a number of comparisons that are relevant in the modern datacenter and that run parallel to the IBM i platform in terms of its transaction processing capabilities.

We endeavor to make more direct comparisons across platforms, provided we can get more data. In the meantime, this competitive analysis from IBM is a good start. Let’s start with transaction processing on relational databases. Here is a comparison pitting the DB2 database running atop the Linux operating system on a Power S1022 against a server using Xeon SP-8380 Platinum processors, which is Intel’s top of the line “Ice Lake” Xeon SP part.

This comparison is not the normal database transaction processing that IBM i shops are used to thinking about. This is one counting number of inference scores per second on roughly equivalent platforms. This particular application has Db2 Advanced Edition V11.5.7 running atop Red Hat Enterprise Linux with the TensorFlow machine learning framework, the LTSM inference model, and a bunch of code running in the Python and Anaconda environments that infers against data extracted from the Db2 database.

What are the big differences between these machines? First of all, the 16-core Power10 system with 256 GB of memory and we presume some flash for local storage costs $84,695, which is 52 percent higher than a 40-core Xeon SP server with 256 GB of memory and we also presume some flash. The IBM machine could do 9,090 scores per second whole the Intel machine could do 9,561 scores per second, which is a 5.2 percent performance advantage to the Intel platform. Aaaah, but wait. Now add in the software and three years of support for the complete system. IBM charges $98,000 per core to license Db2 Advanced Edition, and that drives the vast majority of the cost of both systems. On the Power S1022, the complete stack of hardware, software, and support for three years comes to $1.65 million. The Intel box has 2.5X as many cores, and that really adds up, as you can see, at $3.98 million. (We think it is funny to pay millions of dollars for a single socket machine, which just goes to how you how software costs rise with GDP and inflation and hardware costs go down at a Moore’s Law pace, more or less.)

Here is a comparison that is more akin to the database use case that the IBM i has, and it pits The EnterpriseDB variant of the PostgreSQL open source relational database on a Xeon SP machine and a Power S1022. In this case, the Power S1022 has a pair of Power10 sockets with 16 cores activated using Capacity Upgrade on Demand (CUoD) and has another 16 cores of capacity in reserve that can be activated. This machine is compared to a X86 server that has a pair of Xeon SP-6348 processors with 28 cores each. Both systems have 1 TB of memory and have four virtual machines set up using Red Hat Enterprise Linux 8.6 and its integrated KVM hypervisor. Both machines are running EnterpriseDB Advanced Server 14.3.0 and are running the read-only EDB pgbench. Take a look:

In this case, the 16-core Power10 machine can do nearly 1.23 million transactions per second, which is 19.6 percent more work than the 56-core Xeon SP Gold machine. On the hardware side, this Power10 machine costs $139,095 including the server, operating system, hypervisor, and three years of support, which is 46.9 percent more expensive. But the database software costs $2,000 per core per year, and when you add that EnterpriseDB software and its support to the hardware and operating system costs, the per-core performance advantage of the Power10 chip really shines through. The Power10 system price is $233,175, which is 45 percent lower than the overall price of the Xeon SP system, which weighs in at $423,971. If you do the math on price/performance – meaning cost per transaction not units of work per dollar spent – then one of those pgbench transactions cost just under 19 cents on the Power10 machine, which is 53.8 percent lower than the 41 cents is cost on the X86 machine.

This per-core software and support pricing is obviously a big differentiator, and clearly the cost of enterprise database software is just tremendous compared to the cost of hardware to drive it. So if you feel IBM i with its integrated database is expensive, at least you know other platforms are not cheaper.

This Issue Sponsored By

Table of Contents

Content archive

Recent Posts

Subscribe

Pages

Search