A Cornucopia Of Compute
May 3, 2021 Timothy Prickett Morgan
Ever since the advent of file servers in the 1980s, the rise of client/server system architectures in the early 1990s, and the commercialization of Internet networking in the middle 1990s, AS/400 shops and those using the progeny of that venerable IBM midrange computer have had hybrid computing platforms in the datacenter. Meaning, a mix of processor architectures and operating systems other than OS/400 or IBM i that was in some fashion associated with or actually doing mission critical work.
In fact, as you all well know, there is in aggregate more raw compute in the X86 or RISC servers that sit adjacent to the IBM i platforms than in those IBM i platforms themselves. Sometimes it is multiples like 2X or 3X or 5X, and sometimes it is one or two orders of magnitude. Generally speaking, though, the IBM i systems are running at high utilization doing transaction processing, batch, or data analytics work, while the surrounding layer of machines are doing application serving, file serving, and other workloads but are not running at peak utilization. They have excess capacity and redundancy to make up for some of the shortcomings of the systems, which make them inexpensive and yet underutilized and therefore, when you do the math, expensive in another way.
We are not here to argue about that right now, or to try to convince everybody to move applications off external X86 and RISC machines and onto Power Systems platforms, much less port them to the IBM i operating system. We do this kind of thing all the time. But what we do want to do is make you aware of some of the changes in X86 and Arm server CPU architectures that are either here now or are on the way and how that may impact what workloads run where in your datacenter.
One other thing that we are not going to do right now is go through the nitty gritty of the new “Ice Lake” Xeon SP server CPUs from Intel, the new “Milan” Epyc 7003 server CPUs from AMD, or the “Zeus” Neoverse V1 and “Perseus” Neoverse N2 server designs from Arm Holdings, the company that licenses Arm CPU intellectual property. If you want to see the architectural details of these new processors, then the following links from The Next Platform, where I also happen to work, will be extremely helpful:
- Intel Fields A 10 Nanometer Server Chip That Competes
- Deep Dive Into Intel’s “Ice Lake” Xeon SP Architecture
- The Third Time Charm Of AMD’s Milan Epyc Processors
- Deep Dive Into AMD’s “Milan” Epyc 7003 Architecture
- Arm Puts Some Muscle Into Future Neoverse Server CPU Designs
Intel has been having serious issues with the ramp of its 10 nanometer manufacturing processes in its foundries, and that several year delay has allowed AMD, its archrival for decades in X86 processors, to not only catch up with its Epyc processors, but surpass Intel’s Xeon SPs in many respects such as core count, raw throughput, and price/performance. And the Arm collective, lead now by Ampere Computing and cloud giant Amazon Web Services and a number of smaller players with big aspirations in various parts of the market, will be fielding very good processors starting this year and next. The competition between Intel, AMD, and Arm is going to be fierce, and unlike in prior years, it is going to stay fierce, with an annual cadence of processor updates, transistor manufacturing for these chips being pushed to the limits, and very intense price/performance competition.
There has never been a better time to buy an X86 or RISC server, and we will be able to say that every year going forward from this point forward, until Moore’s Law shrinks of transistors run out and chips start getting a lot hotter and a lot more expensive to make. That day is coming, somewhere between 2025 and 2030. And it won’t be long before what used to be in whole racks of systems will be in water-cooled 3D blocks that look like thermal conduction modules (TCMs) from vintage mainframes.
Given all of this, we offer this advice: “Knowledge is power,” as Francis Bacon famously said over four centuries ago. Get some knowledge and use this as a bargaining lever.
Start a pilot program to study Arm servers. The easiest and cheapest way to do this is to play around with instances based on the Graviton2 processors on the AWS cloud. These are reasonably powerful, and offer very good bang for the buck compared to X86 servers, as we have calculated at The Next Platform. You are stuck with Linux on Arm at this point, however, since Microsoft has only ported Windows Server to Arm chips for its own internal use and does not support Windows Server or the myriad adjunct systems software elements that run with it on Arm processors. And it does not seem inclined to, either. But knowing how Arm server chips work, and how well they work, is going to be a key bargaining tactic, even if you never use Arm servers. And you probably will at some point in the next five years if the price/performance ratios between Arm and X86 server chips hold.
Start a pilot program to study moving X86 workloads to Power. This is important. Do a complete profile of all of the applications you have running on your X86 iron, and see which ones are on Linux and can be moved to logical partitions on Power Systems and which ones are stuck on Windows Server and can’t be moved to Power Systems because Microsoft does not support Windows Server on Power. You would think that IBM and Red Hat would have a concerted, organized porting factory already set up to do this, but as far as we know, the old Migration Factory that Big Blue used to have for moving Solaris and HP-UX workloads to AIX has not been revitalized to move Linux or Windows Server workloads to Power Systems partitions running Linux. So you have to do it yourself.
Make sure you are pitting AMD against Intel when you are looking at buying X86 servers right now. If you have used a single server maker for the past decade or more, break out of that pattern. Get one OEM to pitch systems based on Intel’s Xeon SP processors and another one to pitch systems based on AMD’s Epyc processors. Try to increase the competitive pressure to really drive the deals for X86 iron. You can’t do this by going to one vendor. They will pitch Xeon SPs first, and if you crab a little, they will cut a small deal on AMD Epycs and call it a day.
Buy a lot of capacity now and get a better price for the long haul. We don’t usually advocate for buying more capacity than you need at any time, but as Jensen Huang, the co-founder and chief executive officer at Nvidia has pointed out, “The more you buy, the more you save.” If you make the deal to upgrade X86 servers look bigger by getting a lot more capacity up front, you get a bigger break. There is a time value to money, but there is also a lot less hassle if you just have the server farm to start.
Use aggressive X86 pricing and Arm competitive pressure to drive down Power Systems costs. The competition between Power Systems and other platforms is indirect except for the Linux stack, and that makes it tough to get a good deal. But if you can bring more workloads into the deal, you get IBM’s attention and that of its competitors and they will fight harder to keep the deal. And, if you do it right, you get a more consolidated, easier to use, cheaper platform, including cheaper IBM i platforms, too.
It’s win-win-win. From what you will learn, you will be better positioned to make any move you might need to.