I disagree with the article author and at the end of the super long article, so does IDC. For the big workloads, x86 systems are going to be on top for the next years. Yes ARM is making strides, but it will take time before the move into data centers at any significant rate.
The trend lines from the IDC charts show continued growth, 5-6% growth is pretty decent as you remember manufacturing costs continue to drop.
It appears that he draws most of his conclusions based on what Gartner thinks is going on. Gartner thinks that the next generation of ARM is going to be a huge winner. It take a long time to get traction in the big server marketplace, he thinks it will be sooner.
ARM has some spiffy stuff, but it's not what the big data centers are looking for. There is posts about the Apple CPUs that are out. Nice stuff and if I was building a phone, it would be one of my top choices. But if I'm doing data center then I'm going to want something more general purpose. If I have 6 cores in a chip and they are all the same I can load them up. If I have CPU and GPU cores then I have to move tasks around.
If I move to ARM from Intel then I have the added task of trying to optimize my load/task to the new CPU. Granted compilers are a key driver, but lots of people have spent lots of time hand tweaking Intel CPU loads.
22 cores running at 2.2GHz in a single socket is pretty amazing. Haven't seen any ARM72's anywhere near that big.
You might see some movement in this area, but the way companies are gobbling big servers, we are not at peak Intel.
22 cores running at 2.2GHz in a single socket is pretty amazing.
I'm still trying to wrap my brain around the idea that the processes running on this system aren't going to be memory bandwidth starved. Yes, I know the Xeons have gobs of cache in multiple levels, but still. And yes, I know that distributing tasks across multiple systems can have its own severe performance problems.
But the market keeps getting bigger and bigger. Unless we are using a different version of what "peak" means in this context.
"The hypothetical point in time when the global production of x86 equipment reaches its maximum rate, after which production will gradually decline." is the one that I'm using. As long as the demand is up there, then production will continue to grow.
The Apple A10 is trouncing lower-end Intel kit already. We may be looking at a future of client-side computing dominated by Apple; if they can get their shit in one sock cloud-wise they'll make dents in the server space too. Think Amazon AWS, but running Mac OS X server and fully integrated with OS X. Easy peasy deployment of your Rails app straight from your MBP dev box right to the cloud, with almost zero config management required.
They may have the CPU chops but no way do they have the institutional ability to sell clients and servers to business. They can't even hold onto the educational market which they're losing to Google Chromebooks.
> They may have the CPU chops but no way do they have the institutional ability to sell clients and servers to business.
Proprietary hardware is a hard sale for anyone these days but in Apple's case it's even harder because their previous attempts left many shops with a disinclination to ever use them again. It'd take a serious commitment to forget the era where you had to beg for OS X Server bugfixes by saying they were preventing the purchase of n Macs.
I doubt it. Apple has been systematically withdrawing from the server space for years, and I doubt they are particularly likely to reverse that trend. Servers are relatively low margin with a small brand impact. Apple plays well in consumer electronics where brand has much more of an impact on margin.
Apple buys generic Intel kit for their datacentres, and did this even when making their own server product.
Consuming your own output isn't good for margins, and like many large-scale providers, you must acknowledge vanity hardware isn't any faster than bare motherboards in a rack. The more bare-bones the better.
The XServe and friends were intended for small-scale deployments. They did a few partnerships to build clusters of them for HPC but that was largely a demonstration, not a target customer base.
> We may be looking at a future of client-side computing dominated by Apple
I don't know about that. I recently switched from Apple hardware to chromebook hardware because the chromebook was under $200. I just couldn't justify the extra cost.
The Apple A10 is trouncing lower-end Intel kit already.
The version of this story I saw was based on Geekbench scores. Browsing Geekbench results, I noticed that identical CPUs had scores that varied by huge amounts, as much as 2x. That may be a reflection of a real phenomenon, but it's not what I expect from a "CPU benchmark."
Geekbench also claims to model real-world applications, and I wonder what that means for comparing a phone and a laptop. Obviously there aren't so many apps that run the same way on both platforms. I wouldn't have any trouble believing, for example, that the CPU and video display system on a phone were superbly and efficiently integrated, far more so than in a desktop or laptop, and that would show up in an application benchmark. But I'd be surprised if a phone CPU can really get through many threads of branchy code faster than a big high-consumption part with giant caches can.
Anyway, I'm dubious about the Geekbench thing and I wonder if there are other measurements you have in mind.
Aren't there numerous services that make this scenario just as seamlessly? What would Apple add to it, other than their name? What advantages will Mac OS X server (presumably on ARM) offer over what can be done today?
Why would I as a buyer trust Apple in the server space? They've already withdrawn once and they cannot even keep pace at the workstation level (Mac Pro).
I supposed we can A/UX to the list. So, its A/UX -> AIX -> OS X Server. I do admit an iOS Server would be an interesting product. I guess it would be all containers all the time.
It's not about trusting Apple. It's about recognizing that if Apple can push ARM performance to that level then other companies will eventually catch up as well, and if commodity ARM is actually competitive with commodity Intel then they have a problem.
Intel has a huge competitive advantage here in the bulk of the market, but ARM will find an opening as well. They'll probably coexist for some time.
When this article talks about "Unix systems" as something seperate from X86, what does that mean? Itanium and... SPARC? (Assuming POWER comes under "IBM"?)
Power is a completely separate architecture and product line from "IBM mainframes" (i.e. System z), and "traditional" (i.e. AIX) IBM Power Systems products are generally considered "Unix systems".
(Disclaimer: I work for IBM Power Systems, opinions my own)
I'm not in the space at all, so from outside "Unix systems" seems like a very arbitrary label - since I'm assuming most of the X86 machines are running Linux or BSD which are very much Unix, I guess the "systems" part is where there is some specific definition
Thanks for the correction, I assumed some of the commercial Linux distributions like Red Hat or whatever were Unix certified, but I was wrong http://www.opengroup.org/openbrand/register/
Solaris also runs on x86 though, so a "Unix system" means what? Software licenses? Hardware sold with a software license? But that would also include x86 systems?
It's really a historical label at this point but to a first approximation it means that it didn't start on commodity PC hardware, even if decades later it now runs on x86 boxes.
The ancient history is that mainframes (mostly IBM) were so expensive (upfront + license/support) that many places jumped onto minicomputers starting in the 1960s and that process repeated itself a generation later as hardware continued to get cheaper and (later) Unix went from a curiosity to something a lot of people wanted and was generally both considerably cheaper and — especially appealing – meant you at least in theory weren't just locking yourself into yet another proprietary single-vendor platform.
For awhile, that meant companies like Sun, HP, DEC, Data General, SGI, etc. would sell you their hardware running something Unix-like and this was popular enough that even IBM started selling AIX to avoid losing customers entirely. At some point this was termed “Open Systems Unix” since the theory was that you could mix vendors fairly well, although anyone who worked in that period could tell you that the definition of “Unix” or “C” had significant variations across vendors.
As x86 improved, that cycle repeated again. Unlike its predecessors the 80386 was 32-bit and had hardware memory management and permissions so it was much easier to maintain the Unix model there, and the sky-rocketing volume meant other hardware platforms found it increasingly hard to remain competitive on performance or cost. One by one they either gave up and ported to x86, left the market, or tried to specialize in some niche like SGI's various high-performance computing products with exotic interconnects, graphics hardware, etc.
"Unix systems" is a kind of antiquated term for when the big iron vendors (IBM, Sun, HP, DEC, MIPS) sold large RISC systems with the hardware (Power, SPARC, pa-risc, Alpha, MIPS) as a loss leader for the proprietary UNIX (AIX, Solaris, HPUX, Tru64, RiscOS) running on them. Linux ate most of those proprietary unices, and now the economic logic (for the remaining survivors) is reversed: the money is in the hardware and in support.
So why does AMD bet on Zen instead of creating their next ARM SoC? The A1100 is pretty good on paper except the CPU but they released it two years too late and the availability still isn't great.
According to the article, x86 constitutes 99.3% of the market. I think AMD would rather have a small piece of the 99.3% rather than a large piece of the 0.7%, even if that 0.7% grows dramatically.
Google has been quoted as stating if they could get a 20 percent power/performance advantage buy switching to IBM Power they would do so, even if it were only for a single generation.
My question is, wouldn't this be possible with ARM?
I don't think so. AFAIK, ARM performs well in computation/watt on the low end of computation. But you have to get above a certain performance level to amortize the power overhead of the rest of the computer (motherboard, ram, etc).
This comments thread is off-track. This is not yet another article about how ARM is poised to eat Intel's lunch, and that cites (worthless) Geekbench scores to make that claim. No, this was written by Timothy Prickett Morgan, who actually knows hardware.
His basic argument, at least on my quick reading, is that right now, Intel's Xeon margins are insane -- some 50% -- and x86's market share is north of 99% in servers. Furthermore, Intel, Microsoft, and Red Hat are sucking up all of the margin in the stack, and leaving everyone else to fight for scraps. (We've seen this "Intel eats all the margin" movie before in the PC space, and I wrote about it here: https://techcrunch.com/2016/05/17/how-intel-missed-the-iphon...)
So, these three things -- an x86 monoculture, crazy high margins, and three companies sucking up all the profits from the rest of the stack -- are both individually and collectively unsustainable. Morgan seems to be arguing that a rational market is going to take care of this arrangement one way or the other. By virtue of its margins and market dominance, Xeon has a massive target on its back, and there is a ton of ingenuity gunning for it. So at some point things will change.
So no, this isn't yet another in a near-decade long string of "ARM vs. x86 as told by Geekbench, with a side of OMG Apple" articles. He's actually pointing out that, technical considerations aside, there is a set of market conditions currently prevailing that looks unsustainable on its face, and at some point, like when pressure builds up along a fault line and then gives way, there will be a big shift.
I agree that Intel margins are pretty high. But at 50% margins, when there is a shift, they can move the price point down some.
I just re-read your article and it appears to be mostly about the mid->low power market (ie, Atoms from Intel) as an example.
The Server article posted is about the high end Intel chips, the $4500 ones you find in servers.
Your post is spot on about Intel in general, I think at the low ends we will start to see more shifts away from x86 at the desktop / laptop levels as other chipsets become faster / highly available. But I think in the server space, Intel has lots of running room left in the x86 space.
Morgan is a very smart guy, but I think his reliance on Gartner may have pulled him off the mark.
Drop the "rational" to avoid that baggage, and the argument is certainly valid.
It could be a winner takes all if it was a huge technical hurdle to move many x86-based systems to ARM and back. But it's likely never been easier, and a lot of software already works on x86 and ARM. Unless x86 can consistently perform better for less power consumption and price, there will be competition.
There's aren't any big architectural reasons that x86 would need to be less power efficient. Intel already had a go at that side but very few people bought the low power server chips (Avoton etc). AMD also hasn't had much luck with low power server chips (see SeaMicro etc).
One big reason is that halving the CPU power consumption doesn't nearly halve the server power consumption, but it often goes pretty close to halving the system performance. And CPUs are pretty efficient at idling so you're really only paying for the power consumption when there is work to be done - instead paying for larger fleet of reserve servers.
And of course scaling to more but slower servers means a lot of expensive software/admin complexity.
He's actually pointing out that, technical considerations aside, there is a set of market conditions currently prevailing that looks unsustainable on its face, and at some point, like when pressure builds up along a fault line and then gives way, there will be a big shift.
Fair enough. But why wouldn't the change be something like: Intel buys ailing server businesses cheap, keeps them going with even less margin than before, just to keep x86 flowing into the market? That would end the "unsustainable" situation but wouldn't really shake up the market much.
A long article, but the real quote is "The truly remarkable thing is that the market for servers continues to be diverse. ". But then if you look at the table, which is ranked by revenue you see the ODM market is moving into 3rd place (its behind IBM by accelerating past them) with a 13% growth.
Which is just large sections of the market realizing that paying $800+ for BMC features from HP/Dell that everyone else gives away, or 5x the street price for storage/ram differing only in who's sticker is stuck on top of the micron/WD/etc hardware is dumb. So, the revenue numbers hide the volume difference which is strongly shifing away from the IBM's and to the smaller whitebox manufactures.
The only question is will Dell/HP wake up before supermicro/levnovo finish off what amazon/microsoft have started.
43 comments
[ 2.6 ms ] story [ 102 ms ] threadThe trend lines from the IDC charts show continued growth, 5-6% growth is pretty decent as you remember manufacturing costs continue to drop.
ARM has some spiffy stuff, but it's not what the big data centers are looking for. There is posts about the Apple CPUs that are out. Nice stuff and if I was building a phone, it would be one of my top choices. But if I'm doing data center then I'm going to want something more general purpose. If I have 6 cores in a chip and they are all the same I can load them up. If I have CPU and GPU cores then I have to move tasks around.
If I move to ARM from Intel then I have the added task of trying to optimize my load/task to the new CPU. Granted compilers are a key driver, but lots of people have spent lots of time hand tweaking Intel CPU loads.
22 cores running at 2.2GHz in a single socket is pretty amazing. Haven't seen any ARM72's anywhere near that big.
You might see some movement in this area, but the way companies are gobbling big servers, we are not at peak Intel.
Hope that helps.
I'm still trying to wrap my brain around the idea that the processes running on this system aren't going to be memory bandwidth starved. Yes, I know the Xeons have gobs of cache in multiple levels, but still. And yes, I know that distributing tasks across multiple systems can have its own severe performance problems.
I think the point is that you're pretty much by definition at the peak when your market share is 99.3%. There's nowhere to go.
"The hypothetical point in time when the global production of x86 equipment reaches its maximum rate, after which production will gradually decline." is the one that I'm using. As long as the demand is up there, then production will continue to grow.
Proprietary hardware is a hard sale for anyone these days but in Apple's case it's even harder because their previous attempts left many shops with a disinclination to ever use them again. It'd take a serious commitment to forget the era where you had to beg for OS X Server bugfixes by saying they were preventing the purchase of n Macs.
Consuming your own output isn't good for margins, and like many large-scale providers, you must acknowledge vanity hardware isn't any faster than bare motherboards in a rack. The more bare-bones the better.
The XServe and friends were intended for small-scale deployments. They did a few partnerships to build clusters of them for HPC but that was largely a demonstration, not a target customer base.
I don't know about that. I recently switched from Apple hardware to chromebook hardware because the chromebook was under $200. I just couldn't justify the extra cost.
There's no indication of a switch to ARM-only from any vendor I'm aware of. Lenovo et al. are still releasing new x86-based Chromebooks.
The version of this story I saw was based on Geekbench scores. Browsing Geekbench results, I noticed that identical CPUs had scores that varied by huge amounts, as much as 2x. That may be a reflection of a real phenomenon, but it's not what I expect from a "CPU benchmark."
Geekbench also claims to model real-world applications, and I wonder what that means for comparing a phone and a laptop. Obviously there aren't so many apps that run the same way on both platforms. I wouldn't have any trouble believing, for example, that the CPU and video display system on a phone were superbly and efficiently integrated, far more so than in a desktop or laptop, and that would show up in an application benchmark. But I'd be surprised if a phone CPU can really get through many threads of branchy code faster than a big high-consumption part with giant caches can.
Anyway, I'm dubious about the Geekbench thing and I wonder if there are other measurements you have in mind.
Before the Xserve there was the Apple Network Server https://en.m.wikipedia.org/wiki/Apple_Network_Server
Intel has a huge competitive advantage here in the bulk of the market, but ARM will find an opening as well. They'll probably coexist for some time.
(Disclaimer: I work for IBM Power Systems, opinions my own)
The ancient history is that mainframes (mostly IBM) were so expensive (upfront + license/support) that many places jumped onto minicomputers starting in the 1960s and that process repeated itself a generation later as hardware continued to get cheaper and (later) Unix went from a curiosity to something a lot of people wanted and was generally both considerably cheaper and — especially appealing – meant you at least in theory weren't just locking yourself into yet another proprietary single-vendor platform.
For awhile, that meant companies like Sun, HP, DEC, Data General, SGI, etc. would sell you their hardware running something Unix-like and this was popular enough that even IBM started selling AIX to avoid losing customers entirely. At some point this was termed “Open Systems Unix” since the theory was that you could mix vendors fairly well, although anyone who worked in that period could tell you that the definition of “Unix” or “C” had significant variations across vendors.
As x86 improved, that cycle repeated again. Unlike its predecessors the 80386 was 32-bit and had hardware memory management and permissions so it was much easier to maintain the Unix model there, and the sky-rocketing volume meant other hardware platforms found it increasingly hard to remain competitive on performance or cost. One by one they either gave up and ported to x86, left the market, or tried to specialize in some niche like SGI's various high-performance computing products with exotic interconnects, graphics hardware, etc.
My question is, wouldn't this be possible with ARM?
But that's mostly a guess.
His basic argument, at least on my quick reading, is that right now, Intel's Xeon margins are insane -- some 50% -- and x86's market share is north of 99% in servers. Furthermore, Intel, Microsoft, and Red Hat are sucking up all of the margin in the stack, and leaving everyone else to fight for scraps. (We've seen this "Intel eats all the margin" movie before in the PC space, and I wrote about it here: https://techcrunch.com/2016/05/17/how-intel-missed-the-iphon...)
So, these three things -- an x86 monoculture, crazy high margins, and three companies sucking up all the profits from the rest of the stack -- are both individually and collectively unsustainable. Morgan seems to be arguing that a rational market is going to take care of this arrangement one way or the other. By virtue of its margins and market dominance, Xeon has a massive target on its back, and there is a ton of ingenuity gunning for it. So at some point things will change.
So no, this isn't yet another in a near-decade long string of "ARM vs. x86 as told by Geekbench, with a side of OMG Apple" articles. He's actually pointing out that, technical considerations aside, there is a set of market conditions currently prevailing that looks unsustainable on its face, and at some point, like when pressure builds up along a fault line and then gives way, there will be a big shift.
I just re-read your article and it appears to be mostly about the mid->low power market (ie, Atoms from Intel) as an example.
The Server article posted is about the high end Intel chips, the $4500 ones you find in servers.
Your post is spot on about Intel in general, I think at the low ends we will start to see more shifts away from x86 at the desktop / laptop levels as other chipsets become faster / highly available. But I think in the server space, Intel has lots of running room left in the x86 space.
Morgan is a very smart guy, but I think his reliance on Gartner may have pulled him off the mark.
That's a silly argument. If the market would be rational and it would take care of this we wouldn't have landed in this situation in the first place.
The reality is more that we're more in a winner takes it all situation.
It could be a winner takes all if it was a huge technical hurdle to move many x86-based systems to ARM and back. But it's likely never been easier, and a lot of software already works on x86 and ARM. Unless x86 can consistently perform better for less power consumption and price, there will be competition.
One big reason is that halving the CPU power consumption doesn't nearly halve the server power consumption, but it often goes pretty close to halving the system performance. And CPUs are pretty efficient at idling so you're really only paying for the power consumption when there is work to be done - instead paying for larger fleet of reserve servers. And of course scaling to more but slower servers means a lot of expensive software/admin complexity.
Fair enough. But why wouldn't the change be something like: Intel buys ailing server businesses cheap, keeps them going with even less margin than before, just to keep x86 flowing into the market? That would end the "unsustainable" situation but wouldn't really shake up the market much.
Which is just large sections of the market realizing that paying $800+ for BMC features from HP/Dell that everyone else gives away, or 5x the street price for storage/ram differing only in who's sticker is stuck on top of the micron/WD/etc hardware is dumb. So, the revenue numbers hide the volume difference which is strongly shifing away from the IBM's and to the smaller whitebox manufactures.
The only question is will Dell/HP wake up before supermicro/levnovo finish off what amazon/microsoft have started.