Power11 Entry Machines: The Power S1124 And Power L1124

July 21, 2025 Timothy Prickett Morgan

For several decades, a two-socket server was the workhorse machine in most datacenters, whether they were running a database machine for transaction processing or a scale-out cluster of machines for distributed Web applications and their related back-ends. And for many OS/400 and IBM i shops, the two-socket machine was the one that provided the right balance of compute and storage expansion and density.

But, over the years, as Power processor cores have gotten more powerful, customers have needed fewer and fewer cores to run their relatively modest (by comparison) and relatively static (they grow along with the business, but not faster than the business) transaction processing applications. And so many IBM i customers have one to four cores on Power8, Power9, or Power10 systems.

Those customers who have needed larger four-way and larger systems have similarly been able to downshift to physically smaller machines with fewer cores than you might expect based on past systems. And with the adoption of external SAN storage for these machines, a fairly powerful machine can be crammed into a two-socket server that fits in a 4U rack enclosure.

This is why we chose to start our analysis of the Power11 system lineup with the Power S1124 server. It looks like the workhorse from years gone by, but it has the performance of a big iron machine all crammed into that relatively small form factor.

The front view of the Power S1124.

The Power S1124 system has two sockets, and each socket has a pair of Power11 processors crammed into it that are cross-coupled to each other using NUMA links that are embedded on the Power11 processors, just like that have been on many prior generations of Power iron. (We can’t remember the last time an external chipset for SMP or NUMA clustering was used in a Power server.) The Power11 chip has 16 cores with eight threads per core, just like the Power10 chip did, but with the Power 11, the chips run a little bit faster and the yield is a little bit better thanks to a move to an enhanced 7 nanometer process from IBM’s chip foundry partner for Power and z chips, which of course is Samsung.

In the Power S1124 machine, each Power11 chip has eight, 12, or 15 cores that are activated out of the 16 cores, and the sixteenth core is designated as a hot spare. In the event of a failure of a core, this spare is automagically activated and the PowerVM hypervisor restarts the work that was in progress. Customers do not have to pay for this extra core or for an IBM i license for it.

Customers can buy a single processor module with 16 cores and add another processor module with 16 cores for a total of 32 cores. (That is two different base configurations.) This is feature #EP3X, and the cores run at a base speed of 3.4 GHz and scale up to 4.2 GHz as they dynamically change the clock speeds based on the power draw of the compute complex. Two other configurations come with two processors modules, each with a Power11 DCM installed. One with 24 cores per module runs the Power11 chips at a base 3.05 GHz that scales up to 4.15 GHz and the other with 30 cores per module runs the Power11 chips at a base 2.8 GHz that scales up to 3.95 GHz. Obviously, the per-core performance is going down as the core counts are going up here, but the throughput is rising fast for the system. A machine with 60 cores running at even 2.8 GHz is no slouch, particularly one that has eight threads per core, as Power11 offers. This is called SMT8, and Power8, Power9, and Power10 offered this mode. Power9 offered SMT4 modes in hardware, which had the effect of doubling core counts and making NUMA domains on the processors larger.

Here is what the “Godel” Power11 DCM looks like:

The Power11 Dual Chip Module used in the Power S1124.

The image above was cropped badly by IBM in its Redbook, not by us. This is very hard to read, but the block diagram of the pair of DCMs used in the Power S1024 server would be exactly the same, excepting the DDR5 memory speeds are different and obviously the processors are different in the number of active cores and speeds and thermals and such.

Here is that much clearer block diagram:

On the Power S1124, the differential DIMM memory is based on DDR5 memory running at 4.8 GHz or 4 GHz, depending on the capacity of the memory. The 64 GB and 128 GB memory modules, which are comprised of two D-DIMMs each of 32 GB and 64 GB capacities, respectively, run at the faster 4.8 GHz speed, while the fatter 256 GB and 512 GB memory modules (made of a pair of D-DIMMs of half those capacities) run at the slower 4 GHz speeds. Across two modules in the Power S1124, the 4.8 GHz memory delivers 2,546 GB/sec of bandwidth across 32 memory slots, and the 4 GHz memory delivers 2,048 GB/sec of bandwidth. This is a lot of bandwidth, and because the DDR5 memory is running at much lower speed than the 6.4 GHz memory commonly used in X86 server chips to boost bandwidth across a much smaller number of memory slots and controllers, that extra bandwidth comes with a lower error rate.

If you want the higher bandwidth in the Power S1124, you can’t use the fatter memory, so you top out at 4 TB per machine but you get the extra 25 percent bandwidth (2.5 TB/sec versus 2 TB/sec) in exchange for cutting the system memory in half or to a quarter from the peak capacity using the fatter 256 MB DDR5 modules with OMI front ends on them.

For your reference, here is a table of the salient characteristics of the Power S924, Power S1024, and Power S1124 entry servers from IBM, all in one place:

The Power S1124 does not appear to have OpenCAPI accelerator ports, which is odd considering that eventually IBM will be offering its “Spyre” accelerators for attachment to Power11 systems for boosting AI processing performance by more than is possible using the on-chip Matrix Math Accelerator (MMA) that is on every Power11 core. (And was on every Power10 core, too.)

In announcement letter AD25-0036, which discusses the Power S1124 in detail, IBM says it will offer a Capacity BackUp variant of the Power S1124 system and will also make a Power Solution Edition for Healthcare for the box as well.

The Power S1124 and its Linux-only Power L1124 variant will be available on July 25. We are digging around for hardware pricing for the Power11 machines and how they compare to the Power10 machines and to each other right now. Stay tuned. We presume IBM is giving discounts on the Power L1124 to stimulate Linux adoption, as it has done with L-class machines in the past.

This Issue Sponsored By

Table of Contents

Content archive

Recent Posts

Subscribe

Pages

Search