Power10 Entry Machines: The Power S1022s - IT Jungle

jokbanga.blogspot.com

July 25, 2022 Timothy Prickett Morgan

This is second part of our in-depth coverage of the entry and midrange Power10 machines that Big Blue launched on July 12, and today we are going to take a look at the Power 1022s, a new class of machine in the Power Systems lineup that shoots the gap between the lowest-end Power S1014, which we reviewed last week, and the Power S1024 and Power L1024 that offer more expandability and more features.

We will cover the Power S1024 and Power L1024 next week, and finish off with the Power E1050 – which does not support the IBM i operating system but which could any time IBM felt like it – the week after that.

One thing to note that did not occur to us last week, but does as we think about the new Power10 entry and midrange server line: There is no Power S1012 system, the very entry-est of machine put out by IBM in the past decade that would be a follow-on to the Power S812, an entry machine that was announced in February 2017 to get an even lower price point into the Power Systems line explicitly for IBM i customers.

We called it the Power Mini because, frankly, why be boring? We suggested in late 2017, well ahead of the Power9 launch, that IBM needed to create a machine that we called the Power 911 Mini because a four-core Power Systems machine based on the Power9 processor was going to be overkill for a lot of customers. And, to be fair, a Power 1014 with four cores as a minimum is a lot of oomph for a lot of IBM i shops, and once again we will say there needs to be a Power S1011 with one core and a Power S1012 with two cores to better address the needs of entry customers. And, we also think these machines need more than 64 GB of memory, too, which is where the Power S1014 is capped just like the Power S812, S814, and Power S914 before it.

We are going to think about how it might be best to do this, given the current Power10 system lineup. . . .

The second “s” in the Power S1022s server name is lower case, and presumably is intended to denote a “small” T-shirt size variant of the full-on Power S1022 system; it is going to play havoc with our all-cap titles, of course because it will look like “POWER S1022S.” I never liked those all-cap titles, particularly because it says “IBM I” not “IBM i” but I needed to get onto WordPress quickly and this was an attractive style for a layout at the time in 2018. This may cause me to change it to regular initial caps in the titles. The name “Power S1022s” also means we can’t know if I am talking about plural machines – “the Power 1022s are the most popular machines in the world” – or this machine in particular – “the Power S1022s is the most popular machine in the world” – without paying very close attention. It’s a dyslexia issue, at least for me. And for search engines. A simple dash added like “Power S1022-s” would avoid the confusion.

In any event, this is the effect of the Power S1022s, which was revealed in announcement letter LG22-0031. (I have three and a half decades of six-digit announcement letter numbers, and these eight-digit ones are just a little freaky.) The Power S1014 and the Power S1022s both use what are called single-chip module versions of the Power10 processor, which is supposed to mean one physical Power10 chip is in the socket. The regular Power S1022s, the Power S1024, the Linux-only Power L1022 and Power L1024 versions of these machines, and the Power E1050 midrange system all use dual-chip modules, which cram two physical Power10 processors into a single socket.

But IBM is not exactly using those SCM and DCM terms properly, which is ironic since IBM was among the first to do actual DCMs with the Power5+ chips back in 2005.

If you want to get super technical about it, it looks like there are in fact two Power10 chips in the Power S1022s socket. And I just noticed this now. The second Power10 chip has no cores activated and it is being used as a switch to add extra I/O capability to the first chip. This was also the case in the Power S1014 last week, but it was not obvious in the block diagrams (at least not at the hour I wrote the story). But I am 100 percent sure I am right. Let me explain.

Here is the first Power S1022s block diagram I found, which was given to business partners:

That surely looks like there is a separate and unique I/O chip inside the CPU socket, right? So I took that at face value, and wondered if IBM was getting a PCI-Express switch chip from Broadcom or Microchip and didn’t give it a lot of thought beyond that.

IBM isn’t doing that. If you look carefully, it sure looks like the unique I/O chip is actually a Power10 chip with no cores – and presumably no L1 and L2 cache turned on but maybe L3 cache turned on to act as a funky kind of L4 cache coming off a chunk of the I/O in the box. And that is very clever. IBM gets the benefit of the I/O of the second Power10 chip without paying the thermal price of all of those cores. And it is actually only manufacturing one physical DCM package. Like I said, very clever.

This better block diagram leads me to strongly believe that this I/O chip is a Power10 mostly deactivated:

There are four Power10 chips – P0, P1, P2, and P3 – clearly labeled, and all of the I/O ports are the same type and speed interconnecting the chips. What you have here is a four-way processor complex where half of the CPUs are doing compute and half of the CPUs are only doing I/O.

With the Power S1022s, you can have one or two sockets – the Power S1014 is basically the left-hand side of the Power S1022s in a 4U Power S1024 chassis that can be tipped on its side to make a tower machine.

Here is how the Power S1022s stacks up against the Power S922:

As we have previously explained, the Power10 chip has 16 physical cores and IBM only promised to activate 15 of them in production. As best I can figure, IBM can’t get more than 12 of them active. Rather than have a lot of different core counts as with Power S922 and Power H922 SAP HANA special machines on the Power9 processors, IBM is keeping it simple: the compute part of each socket can have four or eight active cores. Period. The four-core Power10 variant can be used in a single socket, effectively creating a Power S1014, and a pair of eight-core Power10s can be used in the two-socket variant for a maximum of 16 cores in the system. In both cases, the Power10 chips are running at a base frequency of 3 GHz, which can turbo up to 3.9 GHz.

Now, here another twist. The two-socket variant can run IBM i and is in the P10 software tier, but the single-socket variant cannot run IBM i. I am not exactly sure who this four-core variant of the Power S1022s is aimed at – perhaps entry Linux workloads that I have never heard of before? On the two-socket version, customers running IBM i can use PowerVM to carve it up, but there are some restrictions here, too. First, they have to be using the Virtual I/O Server (VIOS) to abstract the storage and network drivers for IBM i, and second, the maximum size of the IBM i partition is restricted to four cores. This sure sounds like the next machine to be deployed in the Power Virtual Server cloud by IBM, and one that it also wants hosters and cloud builders to deploy. Given this, we expect aggressive pricing.

As with the Power S1014 machine, the machine is using IBM’s differential DDR4 DIMM memory technology, which has a buffer chip that speaks the OpenCAPI protocol coming off the Power10 chip on one side and the DDR4 memory interface on the other side. IBM has 3.2 GHz DDIMMs in 16 GB, 32 GB, and 64 GB capacities and across those eight memory channels on a socket it can drive 204 GB/sec of memory bandwidth. The fatter 128 GB DDIMMs coming in the middle of November run at 2.67 GHz and deliver 170 GB/sec of memory bandwidth. Interestingly, there are no 25 Gb/sec OpenCAPI ports in the Power 1022s machine to attach to accelerators or flash or persistent memory.

The Power S1022 machine comes in a rack variant only and is not available as a tower machine optionally. Here is what it looks like:

The machine supports up to eight U.2 NVM-Express flash storage devices, and here are the options:

The NVM-Express drives range in size from 800 GB to 6.4 TB, just like in the Power S1014. Here is what the I/O slots look like, and what you can plug into this 2U chassis:

As you can see, this machine has a fair amount of I/O capability, and does not have the 64 GB memory capping for IBM i like the Power S1014 does. Then again, there is an eight-core minimum because the four-core variant of the Power S1022s does not support IBM i, and for reasons no one has yet to explain but it has to do with driving customers to a P10-class machine unless they can squeeze into that 64 GB memory footprint of the Power S1014.

And if they can, well, that’s amazing. And fortunate. But it sure doesn’t sound right in a modern datacenter. IBM i shops need to be doing more with these machines.

In terms of performance, each four-core partition on the Power S1022s is rated at 106,300 CPWs – the same rating as the four-core Power S1014 running IBM i. If you could get all 16 cores running a single IBM i image, the aggregate compute would seem to be 410,600 CPWs, but it would be less than that due to NUMA and VIOS overhead. But not as much less as you might think given past Power8 and Power9 machines because the NUMA links in the Power10 entry servers have 2X the capacity and lower latency as the links in the Power8 and Power9 servers.

We still do not have pricing, so we can’t do any price/performance analysis. In due time. Up next, the Power S1022 and its Power L1022 variant.