ARM server hype ramps faster than ARM server chips If I didn't have to man El Reg's systems desk for a paycheck and had a little venture capital to blow, I might start a company called Leg Systems, headquartered on the Isle of Man – not because of its tax haven status (which is eroding), but because my company would sell ARM-based systems and say that we wouldn't charge an arm …
Surely this level of competition is good?
AMD vs Intel has sort of died off a bit. Remember Cyrix too? oh and Transmeta. Intel killed them all dead and it was competition from AMD that gave Intel a good kick up the backside and forced them to produce proper multicore CPUs not two P4s stuck together.
@AC,
"Surely this level of competition is good?"
It certainly is!
Intel are in the process of being caught with its trousers round its ankles. The process is quite slow, and because Intel is used to innovating at a pace sufficient to outstrip AMD (and the others you mentioned) they've not seen ARM coming along.
My own view is that several things are happening all at once:
1) The absolute requirement for x86 binary compatibility has evaporated in the space of a single year (about four years ago).
The mobile revolution has shown that the software world is more than happy to abandon x86 binary compatibility. More importantly the Linux world simply laughs at the mere idea of it; most stuff compiles / runs just fine no matter what the underlying CPU architecture is (all other factors being equal).
2) The world has decided that power consumption really does matter. Running a big data centre takes a lot of juice, and that costs a lot of money now.
The mobile world has shown that ARM is the way to go to get good power consumption. Intel's problem is that it takes a lot more transistors (pipelines, instruction re-coding, enormous caches, speculative execution, etc) to make the X86 instruction set run well than are needed for ARM. Transistors consume power. It means that X86 has a built in disadvantage, with Intel staying 'ahead' only because they're really good at silicon manufacturing.
3) Everyone's realised that Intel's "Do everything in the CPU" is wrong.
ARM have comprehensively shown that Intel's philosophy is wrong. Most of the compute intensive things that your mobile does (e.g. video decoding) runs on dedicated co-processor cores or the GPU, not on the ARM core itself. That saves a bunch of power. The same will happen in ARM server chips I expect.
So, what can Intel do?
There's not really a lot Intel can do about this now. The time to have acted was about 7 years ago. To really compete they'll have to go to an entirely different architectural design. That didn't work so well last time they tried. Itanium - remember that? Ironically, that failed because of a perceived need for binary compatibility. It might have succeeded if they had done a really cheap licensing deal with AMD long before launch.....
They could risk building a whole new architecture, but it could be difficult to persuade the software world to follow.
They could re-start building ARM devices themselves but that means following ARM, not leading them. However that's their best bet I think. Intel's silicon mastery could make Intel ARM chips much better than anyone else's, which would mean good profit.
Also they're not helped by the fact that 'ARM' doesn't really mean one single thing anyway. Watching Intel trying to make X86 compete with ARM is a bit like watching an elephant trying to swat a swarm of mosquitoes. Nowadays the mozzies' teeth are getting bigger...
To go even further down the memory lane, Intel's other attempt at a new ISA was the i860. Readers of the grey-haired persuasion will remember articles in Byte [yes, the paper edition] and this chip appearing on boards, possibly alongside an i486 or used on the graphics add-on for the NeXT cube.
@Destroy All Monsters,
I remember the i860 well, even fiddled around with them for some embedded signal processing. They were pretty difficult to get peak performance out of...
The story I heard recently was that there was a single designer in Intel who was largely responsible for that chip. When he left there was no one left who really knew anything about it, so they dropped it.
Very. They all use the same instruction set, so you only need one compiler and any application that runs on one of those CPUs will run on all of them. OSes obviously will need to be ported to each CPU as the interconnect and other IO will be different.
ARM CPUs are more compatible than x86 CPUs as ARM controls the architecture specification, so you don't get different companies producing different, incompatible instruction set extensions like AMD and Intel.
Err, that's more than a little bit optimistic.
There certainly are different versions of the ARM instruction set out there, and not everything that will run on ARMv5 will also run on v4. There are also a whole load of differing extensions in areas like floating-point. So you end up with software distros like debian supporting the lowest common denominator and losing performance because of it, and other distros like raspbian (for the raspberry pi) that are tuned to single devices or chip categories.
ARM is definitely more of a heterogeneous than x86. And yes, there's not usually a device-discovery system or BIOS equivalent, so the OS needs code for the specific board its running on.
If you are a computer maker. Microsoft expects you to include a version of Windows on your systems as well as paying a license fee. But there is no version of Windows available (yet?), so server makers will just forgo the licensing and ship servers with (god forbid) a non windows operating system.
Oh, the horror.
On another front... Apple announces a change in processors for their computer line. Sorry, Intel! Pigs do fly! Pigs in space!
So many CPU designs and configs. Can you imagine the horror of the MS license plan for that?
That said, I'm struggling with the use case for ARM in the datacentre. ARM is good at doing very little tasks which is what we try to avoid in datacentres. As much as I would love to see Intel get a bit of a kicking, I'm not (based on what I see now) that this is it.
ARM corporate desktops, now that might be interesting because they do spend a lot of time idle. ARM ultrabooks (sorry, couldn't resist the irony) also might work. ARM home servers, also good.
Where things might get interesting is with Samsung and HP, where chip-to-retail ownership might yield greater profits over non-vertically integrated intel.
It might also be interesting to see some sort of VM migration with Atom and i7/Xeon in the same box (same cpu package?) When the workload is light, migrate processes to the atom and power down the big cpu, kind of like, well, ARM.
...but that's precisely what WinRT is for!... Joe punter'll get a tiny little useless RT turd for their $50 Windows tax contribution... and gratefully skulk off to discover how utterly crap, useless and inexplicably over-hyped WinRT^HHHHH ARM rubbish is.... then go begging back to the dealer, a month's wages in hand, for another hit of Wintel crack.
Bollox, I'm a Linux only user for the past 10+ years but this Christmas I bought a Dell XPS10 - my first windows since NT4.
I'm quite impressed with WinRT, the responsiveness is excellent, switching apps - and I launched a ton of them - was very fast. The only thing that annoyed me was the network sharing. Yes I could manually plug into to a samba share via //servername/sharename but it didn't integrate as well as expected. It wants to use something called a homegroup. I didn't bother setting it up as I couldn't be arsed.
Sorry, but I may have to down vote, as you have disobeyed the Reg Forum rules.
Rule 1. Though shalt only comment on devices you have never used.
Rule 2.Though shalt only comment on devices you have never used. you have an irrational hatred of, based purely on what you have read on other forums, based in fact or not. (ideally not)
Rule 3. Though shalt never post postive comments. If the device is good, you still must slag off an alternative device in order maintain the bile fuelled hatred on El reg.
This post has been deleted by its author
To me ARM could be successful in the server space in the same way thin clients replaced desktop PCs and Linux won the desktop war over Windows.
ARM enjoys a huge success in the end user device space because:
- End-users don't care about processors so long as they can play angry birds
- ARM's quality (low power consumption) is accentuated on small devices while its disadvantages (bad performance, bugs and incompatibilities) are well hidden by experiences already limited by slow internet connection speeds and latencies as well as small form factors.
- From the manufacturer point of view it limits dependency on the market leader Intel and enables full customization to optimize experience to form factor, e.g. with graphics gpu integration. Product lifecycles are so short that they mostly hide bugs and security vulnerabilities brought by customized silicon;
These advantages will mostly turn to disadvantages in the server space:
- End-users, in this case corporation, will care about the processor, its performance, compatibility and reliability more than performance per watt. And it remains to be seen that ARM has a performance per watt advantage over Intel offerings for server applications;
- End-users may not want to pay increased cost of heterogeneous server software licenses, services and maintenance;
- Also corporation may not want the processor to be produced by Chinese or Korean fabs;
- Server manufacturers will care less about the customizability of the processor because they know it will be a costly nightmare to fix bugs and security vulnerabilities in the silicon they produce; they will care less about all the linux kernel writing they would have to contribute and maintain to ensure adequate performance
- With much longer product lifecycle it will be harder to hide inevitable bugs and security vulnerabilities that invariably increase with multiple code/design forks;
- CPU interconnects are very hard to design with standard homogenous processors; It will be even harder on heterogeneous ARM processors;
- Assuming there is a single ARM design that “wins” and that all server ARM are the same, i.e. don’t suffer from heterogeneousness what’s the advantage over an Intel chip when Intel has arguably the best fabs in the business? Why wait for TSMC to produce server processors 2 years late on a fab process 2 cycles late?
So while ARM is clearly a potential threat to Intel’s server business, and Intel will have to improve its offering and probably lower prices, ARM supporters will face tremendous challenges in this space.
In addition Intel is already finding counter measures in the device spaces, with its latest atom arguably better than a Tegra 3 at lower power consumption. Intel will face the tremendous challenge of convincing phone and tablet OEMs to adopt its platform, on the other hand some manufacturers may not like the ARM dominance of players like Samsung (most probably already a critical supplier of other key components such as memory or screens) or the dependency on not always reliable fabs.
@ Sil,
"These advantages will mostly turn to disadvantages in the server space:
"End-users, in this case corporation, will care about the processor, its performance, compatibility and reliability more than performance per watt. And it remains to be seen that ARM has a performance per watt advantage over Intel offerings for server applications;"
I'm afraid I disagree - performance per watt is very important to the large players (Facebook, Twitter, Google, etc). There are many challenges ahead in getting ARM into the server space in a big way. However the really, really big driver is power consumption.
For the really big data centres it is the dominating cost; more than manpower, more than equipment costs, more than anything else. If they can make big inroads on power consumption by spending a bit more on hardware, design, staff, software, security, maintenance, etc. then they will do it. Their shareholders' demands will make them do it. And if someone makes that fairly easy for them (e.g. Calxeda) then that provider would be in a position to push their wares to the smaller data centre operators too, all of whom would like to save costs too.
Most of what Intel has done regarding power consumption has been based on improved transistor manufacturing, not fundamental architectural innovation. The X86 architecture with it's instruction translation, pipelines, caches, branch prediction, speculative execution, etc. is at a disadvantage compared to ARM which need far less of that sort of thing. And less, so the saying goes, is more.
I think Intel's best bet would be to embrace ARM/server wholesale, and keep doing X86 too.
> Also corporation may not want the processor to be produced by Chinese or Korean fabs;
Fact: Intel has fabrication plants in China [Dalian], also Israel [Qiryat Gat], so I don't see where there is a problem. All their Assembly Test Facilities are in China, Vietnam and Malaysia.
I hear about all these ARM servers available, but where can you buy them? I have an OpenRD box which is pretty good, but it has a fairly old CPU by today's standards...
A modern replacement, rack mountable with 2x GigE, onboard SATA and a PCIE slot would be great but all i see are either cheap development boards which are too limited, or highend boxes which try to pack hundreds of cores into a small space which are huge overkill for my needs and have very high startup costs if you want to start small and expand later (eg having to buy a blade chassis)...
If i could buy a 1U ARM server with the above specs i'd be all over it, one for testing to start with and more later if they worked out ok.
Due to ARM's licensing model I predict that there will be more fabless chip makers creating processors for a myriad of special applications. It may even that corperations will be able to design and aquire processors and motherboards customized for their needs, for which their own operating system and applications can be created and compiled. Linux is like Lego, it's mix and match, why shouldn't this model be extended to the hardware? All we would then need is a 3D printer or similar to create this motherboard or that PCI card, metal sheets and a punch press for the server case, fans, power supply and a bunch of cables....done.
But why stopping at the CPU, why not extending the chip licensing model to RAM memory, SSD storage, etc.?
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
