I guess x86 is king, but what other CPU architectures are commonly found in data centers?
There are still some Sun (Now Oracle) Sparc machines around.
IBM RISC machines running AIX. Itanium running VMS. Not sure if they would be called common compared to x86.
Lots of them, in fact.
IBM servers have their own architectures and chips, and there are still Itaniums.
IBM Power systems. Next thing will be PureScale.
ARM processors are used a lot for servers. They are a bit brain dead compared to x86 processors, especially with respect to floating point math, but data centers don’t need that. The advantage of ARMs is that they are low cost and low power. Cheaper servers and cheaper electricity costs get very significant when you have an entire room filled with racks and racks of servers.
MIPS and SPARC come to mind as well.
The PowerPC (now called Power Architecture) has dropped out of the home computer market and is now targeting servers and embedded applications.
On this site, you can see the architectures of the top 500 supercomputers. I know that is not exactly what you are asking, but it is an interesting data point.
In big iron data centers you are going to find some of x86, Itanium, Power (IBM), Sparc (sun) and some others (Not necessarily in that order. Others may give better info on that.).
If by “servers” you mean servers that are in many businesses of 500 employees or less, I am sure x86 will outstrip the others by far (mail servers, file servers, web servers, etc).
Big iron jockey here. While I may have a somewhat obvious interest in mainframes, I’d wager the number of ARCHLVL2 (System z9) or ARCHLVL3 (System z10) machines running z/OS is pretty tiny compared to ones running x86, Itanium, PowerPC or even SPARC.
We might have ten z10 machines, but we have nearly 4,000 Sun SPARC boxes, over 13,000 Linux boxes and roughy 22,000 Windows servers. Most companies don’t have any mainframes, but most large enterprises have at least one mainframe (or at least an AS/400 because those things refuse to die) running their core business processes such as real-time inventory and point-of-sale.
As a rough breakdown for us, that’s 35,000 x86 or Itanium boxes, 4,000 SPARC, and 10 z10.
In my company’s case, we turned off the last PA-RISC based boxes this year. SPARC went out some years ago and now we have only x86 left.
Incidentally, NASA recently announced that they retired their last mainframe.
x86=70% of server revenue
Z, Power, Itanium, Sparc=30% of server revenue
x86 dominates at the low/medium end (and constantly growing upward) but the architecture doesn’t scale well enough to take over the high end, yet.
Z, Power, Itanium, Sparc are common at the high end (and constantly retreating).
What do you mean by “data centers”? Ones that are built / operated by one business for their own internal use, or ones where anyone can rent space and install whatever type of system they want, or something in between?
A system that serves web content / streaming video / etc. is going to need lots of storage, not a massive amount of memory, and not need a top-of-the-line CPU. Probably the most CPU-intensive thing a box like that will be doing is the encryption of https content. And there are dedicated accelerator cards for that.
There are far more of those systems out there than there are systems doing weather prediction or analyzing data from physics experiments.
In my personal data center, I have 2 Dell servers and 3 custom-built servers (all Intel x86), and one DEC/HP Alpha system. There are 4 Windows desktops (all Dell, Intel x86) and 3 Windows laptops (2 Dell Intel x86, 1 HP AMD x86).
In the shared data centers where I have equipment, I’d say the overall population among all occupants was at least 85% x86. I see the occasional Alpha system and some Sun hardware.
I work for a company that is a direct competitor with IBM, HP and CA. We have about 600 products on every platform I can think of. Here is where our sales are:
z/OS - This is still a mainstay. We estimate only a few thousand companies are still using these. And there is more or less zero growth here. We have about 90% market penetration with at least one product. This is an important, but relatively speaking, static market.
x86 - This is now our core processor. Between windows, the most common linux ports, and the fact that many of the other OSes (Solaris, BSD, OSX) have ports to this architecture, most of our new sales are here.
Power - This is where IBM lives. Both P-series (AIX unix) and I-series (AS/400) appear to be growing markets. AIX is apparently the most widely used unix now. AS/400 is still niche, but unlike mainframe, does appear to be growing.
Itanium - This is where HP lives. HP-UX, VMS, & NonStop Kernel are all exclusively Itanium now. But the older chip-sets are still maintained and supported (PA-RISC, Alpha, & MIPS). None of these OSes are seeing much in the way of growth though. And over the past few years the other venders who have supported Itanium have discontinued that support. Window, SuSe, and RedHat have all left the architecture… leaving HP as the only vendor.
SPARC - This is where most Solaris installs live. Right now I think Oracle is still selling more Solaris on SPARC than on x86. But the x86 support seems to be getting much better and I get the impression that is where Oracle intends to go.
I don’t have any hard numbers, and the ones that are released tend to be highly debatable. But I think RaftPeople has it about right. If you look at sales figures, Something like 75% of server are x86 and most of those are Windows. However when attempts to identify actual servers by polling them (generally internet facing devices) Windows market share drops to about 35% of the market, with Linux and Unix dominating the rest. One recent survey showed this:
Windows - 36.5%
UNIX - 32.7%
Linux - 13.4%
z/OS - 11.8%
The z/OS share looks way too big to me. But otherwise it looks about right to me. x86 dominates. As both Windows and most of the Linuxes will be on x86. All the others combined account for less than half the market.
So “real work” is not serving up a search of the Internet in less than a second from anywhere in the world, or streaming any one of a hundred million videos on-demand almost instantaneously, it’s running a mail merge and printing out “You have been selected for a free vacation to this lovely resort”,
No, real work is screwing up 300 million accounts in less than 30 seconds because someone ran the wrong job.
Yeah, been there, got the 2:30 AM page. The transactional throughput of a mainframe can be frightening, especially if you’re the operator who just ran some JCL cards out of order, disabling every ATM card belonging to account holders with a vowel in their last name. Not that this has ever happened, but a mainframe would do it very quickly.
Not too surprised that NASA is no longer running any mainframes as a Beowulf cluster can probably crunch numbers about as well as a mainframe for less cost, but it’s not all about the gigaflops in banking, which is where I’m at. Although, a z9 is not exactly old and tired. They probably just had trouble finding people who know FORTRAN.
Literally “cards”? I haven’t worked on mainframes in 30 years, but it was cards back then. Are 3270 terminals still around? What languages are used on mainframes these days? I remember the old joke about the new object oriented COBOL called “Add one to COBOL”.
No physical cards any more, but the name stuck, just like dialing a phone.
As for 3270 terminals, we finally got rid of the last real one about seven years ago in favor of terminal emulator apps. It was being maintained as an “emergency” backup in case the network went down, but we dumped it when we discovered someone threw an Ethernet converter on its line in the wiring closet, so the redundancy was an illusion.