Power10 Entry Machines: The Power S1024 And Power L1024
August 8, 2022 Timothy Prickett Morgan
This is fourth part of our in-depth coverage of the entry and midrange Power10 machines that were announced on July 12 and that started shipping a few weeks later. This week, we go over the last of the machines that can run the IBM i operating system, the Power S1024 and its quasi-Linux-only companion, the Power L1024.
We already went through the Power S1014, the Power S1022s, and the Power S1022 and its quasi-Linux-only variant, the Power L1022. And next week, we will take a look at the midrange Power E1050 – which does not support the IBM i operating system but could if IBM changes its mind. Then, provided we can get pricing information, we will dive into performance and price/performance analysis to help you compare these Power10 machines to prior generations of Power Systems entry and midrange iron.
The Power S1024 and Power L1024 was revealed in announcement letter LG22-0032 and its companion Power L1024 was unveiled in announcement letter LG22-0033. On the L-class machines, Linux is the primary operating system and a maximum of 25 percent of the cores can be activated to run either IBM i or AIX. As with the other Power10 entry and midrange machines announced on July 12, the Power S1024 and Power L1024 were both available on July 22, with a 256 GB memory card feature that gets these machines to their maximum 8 TB of memory shipping on November 14.
And just for the record, 8 TB of memory for a two-socket machine is huge, and that is no accident. These machines have the memory capacity and memory bandwidth to take on some pretty hefty workloads, including in-memory databases and various kinds of analytics. And in some cases, we think, the premium that IBM will charge for a Power10 entry server will be justified compared to Intel Xeon SP or AMD Epyc server because it has 2X the memory capacity and somewhere between 3X and 4X the memory bandwidth. And not just for the SAP HANA in-memory database that has been a big driver of four-socket, eight-socket, and sixteen-socket Power8 and Power9 iron in the past seven years.
The naming conventions of the Power Systems machines tell you a bit about what the machines are, which is a kindness as well as logical. The “S” tells you it is an entry machine and the “L” tells you it is a Linux-only (well, a “Linux-mostly” entry machine), the “10” tells you it has a Power10 processor, the “2” tells you it has two processor sockets, and the second “4” tells you it comes in a 4U rack-mounted form factor.
Here is how the Power S1024 and Power L1024 rack-mounted machines stack up against their Power S924 and Power H924 predecessors:
As we reported with the 1022-class machines, IBM has done away with the distinction between the SAP HANA H series and the Linux-mostly L series machines, which stands to reason since HANA requires Linux and IBM already had a “price-optimized” Linux-ish pricing scheme to better compete with X86 servers – mostly those based on Intel’s Xeon SPs, but sometimes those based on AMD Epyc CPUs.
The Power S1024 and Power L1024 machines have the same salient characteristics, so to keep things simpler, we are just going to say Power S1024 as we go through this. And for the Power S924 and Power H924, we will similarly just say Power S924 in the comparisons that follow.
The Power S924 had single-chip Power9 modules (SCMs) that came with 8, 10, 11 or 12 Power9 cores activated per socket. The Power9 chip topped out at 12 cores in SMT-8 mode, which is with eight threads per core activated. The Power S1024 has a dual chip module (DCM) processor card that has a total of 32 physical cores and 12, 16, or 24 of those cores can be activated. The other cores on the Power10 DCM do not work, and it is common in the CPU industry to pack extra cores in the design to increase the effective yield of a chip. All CPU and GPU makers do this, but we are surprised that IBM doesn’t have full core count options available, or something closer to the maximum. The best IBM can do is 75 percent of the cores right now, and the more commonly shipped SKUs will be at 37.5 percent or 50 percent. The Power S1024 supports up to two processor cards, and thus the system can have 24, 32, or 48 cores all sharing the same memory space.
This will be an enormous amount of processing capacity for most IBM i customers. In the Power S1024, feature #EPGM has 12 cores running at a base speed of 3.4 GHz with a turbo speed of 4 GHz. One of these cards with 12 cores is rated at 312,500 on the Commercial Performance Workload (CPW) benchmark test that IBM uses to rate the relative performance of OS/400 and IBM i machines. With two sockets and a total of 24 cores, the machine is rated at 585,100 CPWs, and that means there is only a 6.4 percent overhead in making the NUMA connections between the four physical processors in the Power S1024 machine. That’s pretty good, and is no doubt driven by the higher performance of the NUMA links between the chips in the DCMs and across the DCMs.
Feature #EPCG has 16 cores running at 3.1 GHz and it can turbo up to 4 GHz as well. With 32 cores fired up across two processor cards, this machine is rated at 725,000 CPWs. And feature EPGD has 24 cores running at 2.75 GHz with a turbo up to 3.9 GHz. With the full complement of available cores spinning, a two socket Power S1024 is rated at 947,500 CPWs. That is more than three times the performance as a 32-socket, 64-core Power 595 based on dual-core Power5 processors running at 5 GHz from 2008 and almost as much oomph as a Power 795 with 192 cores running at 3.7 GHz from 2010 that was rated at 1.19 million CPWs. The Power E1080, you will recall from our price/performance analysis last fall, is rated at 5.27 million CPWs with 240 Power10 cores running at 3.55 GHz.
The Power S1024 has up to 32 directional DDR4 memory DIMMs per machine, and a total capacity of 8 TB with the fattest sticks in them running at 2.67 GHz. With the skinnier sticks running at 3.2 GHz, the machines have a maximum memory bandwidth of 818 GB/sec, which is the kind of memory bandwidth you see in many GPU accelerators, not CPUs. And this is without resorting to HBM stacked DRAM, mind you, as Intel is going to do with its Xeon SPs and that GPUs use to crank up the bandwidth. As for I/O, there are a total of 128 lanes of PCI-Express, which can run at a mix of PCI-Express 4.0 and PCI-Express 5.0 speeds, as shown in the table above.
Like all of the other Power10 entry machines, the Power S1022 only supports NVM-Express flash storage – no SAS or SATA flash and no SAS or SATA disk drives. This is the 21st century. . . . There is plenty of I/O bandwidth in the Power S1022 to balance it all out.
Here is the layout of the system board on the Power S1024:
All of the four “Cirrus” Power10 chips in this machine are meant for compute and, unlike the Power S1014 and Power S1022s, the second Power10 chip in the DCM does not have all of its cores turned off and it is not being used as an I/O switch for the first Power10 chip that does have compute.
Here is a much better block diagram of the Power S1024 system board:
Like the Power S1022 and its Power L1022 variant, the Power S1024 and its Power L1024 variant is really a four-way NUMA server masquerading as a two-socket server. The sockets are arbitrary boundaries to a certain extent. The 1024-class machines have more peripheral expansion than the 1022-class machines, and that is enabled by having a chassis that is twice as high and making good use of that space.
Like the Power S1022s and the Power S1022 and its Power L1022 variant only come in rack-mounted form factors, which look like this:
The 2U form factor supports a maximum of 16 NVM-Express flash drives.
Just like the other Power10 entry machines, the Power S1024 can have NVM-Express flash drives in 800 GB, 1.6 TB, 3.2 TB, and 6.4 TB capacities for a maximum of 100.4 TB of local storage capacity.
Here is how the PCI-Express I/O lays out in the back of the Power S1024 machine:
This is as much I/O capacity as many IBM i customers will ever need, particularly given the performance of NVM-Express flash at this point. It’s just amazing what can be packed into a 4U chassis these days.
Up next, the Power E1050, and why we still want it to run IBM i.