ARM: The nature of servers has changed

ARM: The nature of servers has changed

Summary: Over the past few years a major shift has occurred in servers and the types of workloads they run, and this shift could bring ARM's chips into large cloud datacentres, threatening Intel's dominance.

SHARE:
TOPICS: Cloud, Servers, ARM
22

Changes brought about by the rise in cloud computing have had a huge impact on servers and server design, and could be the key factor in bringing ARM's low-power RISC chips into the datacentre.

ARM believes that the rise of cloud companies like Facebook, Google and Amazon is bringing about a change in attitudes to processors that could make its chips relevant for these companies, posing a potential threat to Intel and AMD's server businesses.

Over the past few years, "the nature of servers has changed," ARM's general manager of its processor and physical IP divisions, Simon Segars, told ZDNet this week. "It's really the growth in cloud [and] growth in companies doing web hosting [and] social sites. As that has grown the nature of servers has changed."

These workloads place the emphasis on light computing tasks but with lots of parallelism and frequent use of large amounts of data, Segars said. Because of this, people are becoming ever more concerned with the power consumed by each individual processor, and this could bring ARM chips into the datacentre.

Companies want the thermal design power - how much power, roughly, a chip uses - to be "as low as possible," Segars says. ARM's chips, which sit at the heart of the vast majority of the world's mobile phones, including Apple's just-released iPhone 5, consume much less power than Intel's processors.

Calxeda's EnergyCore ARM server consumes around 5W, and the company published the results of an ApacheBench benchmark in June that showed its technology beating a 102W Intel Xeon processor. 

Over the past few years, ARM has been developing new chips that are designed for servers as well. In October last year, it announced a 64-bit design. 64-bit is considered to be a must-have feature for servers, so in about a year when ARM licensees start churning out the processors, big things could happen in the datacentre.

This change will be driven by the "mega-trend" of the growth in data-intensive mobile devices, like smartphones, Segars said. "A lot of those data services can be streamed by ARM-based servers."

ARM's challenges

The counter argument to ARM's is that Intel has years of experience of building server processors and has developed a large amount of x86-specific technologies to support server workloads.

Another one is that as ARM designs new chips for the datacentre it will have to put additional features into the processors, which will lead to a rise in the amount of power consumed.

Meanwhile, Intel is preparing to launch its Haswell processor, which will consume as little as 8W of power, taking a high-end compute chip into territory typically occupied by ARM.

Furthermore, ARM has less broad support for the types of software that run on servers, though Segars notes this is changing: "In the server space the Linux kernel has been optimised for ARM for many years," he said.

For the small and medium business there is no evidence that ARM chips have a credible play there, with these companies typically preferring x86 servers from major enterprises like HP, IBM and Dell due to a combination of the Intel processors, comprehensive service and support contracts, and the related software ecosystem. There is not really an equivalent ecosystem of software for ARM yet.

That said, these vendors are interested in ARM themselves. HP has produced a prototype ARM-based server using Calxeda's technology as part of its Project Moonshot hardware development scheme. 

ARM gambles on success in the major clouds

ARM's bet, though, is that the very large cloud operators - Google, Amazon, Facebook - could be tempted by its chips due to the impact it will have on their datacentre electricity bills. Over the lifetime of a datacentre, it's typical that at least 50 percent of the cost of the facility comes from the electricity it uses, so driving this down is a priority for these companies.

Related to this is the rise of software-as-a-service tools for small businesses like Salesforce for CRM or Microsoft Office 365 for productivity. Adoption of these tools means businesses have less need for hardware themselves and instead outsource the capital cost to a cloud operator.

"If you're focused on being a dental practise of a hospital, the last thing you want to do is manage IT," Ian Ferguson, ARM's director of server systems and ecosystem, told ZDNet. "Fundamentally those guys aren't going to be buying as much IT equipment in the next few years."

Whether ARM's chips have what it takes will become apparent over the next couple of years as 64-bit designs come onto the market.

"I think you can regard ARM in the cloud as a player in 2014," Ferguson said. "There are trials going on in end users right now. I think those will start to translate to volume in that sort of timeframe... it takes some time to get the silicon right, to get the platforms right, and on 64-bit we've got to get more of an ecosystem going."

Topics: Cloud, Servers, ARM

Jack Clark

About Jack Clark

Currently a reporter for ZDNet UK, I previously worked as a technology researcher and reporter for a London-based news agency.

Kick off your day with ZDNet's daily email newsletter. It's the freshest tech news and opinion, served hot. Get it.

Talkback

22 comments
Log in or register to join the discussion
  • i wish people would stop spamming

    Their own sites and discuss the articles at hand...

    ARM makes good low-power CPUs but, on an equal benchmark, ARM might use less power but Intel would offer more CPU horsepower... something busy datacenters need too...
    HypnoToad72
    • data centers

      are generally more concerned about their power bill than squeezing a little extra juice out of each chip
      theoilman
      • per watt

        its not about CPU power, its about CPU power per watt. If Intel makes a CPU that does 100 transactions per second (or whatever) at a cost of 100W, and ARM makes a CPU that does only 50 transactions per second, but at a cost of 40 watts... then ARM is a better option - obviously you'll need to buy more servers, or put more CPUs in each server, but in your datacentre you're already running loads of servers anyway and space is cheap.
        gbjbaanb2
  • 2014? by 2014 haswell will be old news.

    Intel will be shipping 14nm parts in volume that use dramatically less power than haswell and at probably 20x the perf of arm. If you're going to do any capex for your datacenter in 2014 Intel will be a much better choice than arm. Note Intel will also be the best choice for tablet and even smartphone as well. They'll have arms butt kicked up and down the compute spectrum. And that's before their move to 10 nm and below where arm probably will trail them for a decade. Remember once Intel grabs the lions share of tablet and smartphone arms revenue is going to dry up pretty fast and they won't even be able to join together enough money for future foundry overhauls.
    Johnny Vegas
    • and you seriously think

      that ARM is going to sit around and do nothing? ARM designs are moving about as fast as intel designs, there's no reason they can't keep up in general and keep ahead on power consumption.
      theoilman
    • ARM don't carry foundry cost

      Hey Johnny
      Quick point (about to get on a plane). ARM don't operate their own foundries, though they invest a bit in physical IP to qualify chips on them. Bulk of foundry cost carried by TSMC/GloFlo/Samsung, et all. I recommend you have a read about ARM's licensee model (if you aren't already familiar with it) as it gives them a few advantages in their competition with Intel.
      Thanks for commenting!
      Jack
      Jack Clark
  • ARM: The nature of servers has changed - Software Licensing

    Software licensing will take 2-3 years to catch-up to the ARM servers. There are way too many oddball metrics for server licensing to make the ARM servers good anywhere outside the cloud and Open Source licenses. Right now you can play the Intel versus AMD processor game because some SW manufacturers have priced their titles so one or the other is a good way to go. Those same SW Mfgs will need to spend a good deal of time figuring out how to price the ARM processors into their titles and make them competitive.
    oien@...
    • Re: Software licensing will take 2-3 years to catch-up to the ARM servers.

      Sounds like you’re talking about Windows Server, which nobody wants to run on ARM anyway.
      ldo17
      • ARM: The nature of servers has changed - Software Licensing

        Doesn't make any difference what OS. Eventually you get to the point you need to run useful software at an Enterprise level. You need tools to manage the server, the OS even paid support for Linux), as well as the apps. I have seen the 40' containers with thousands of servers running - 4 servers per 1U. I get it. My point is the HW will be way ahead of the SW when it comes to licensing - we won't see this as a data center play until licensing cathches up.
        oien@...
        • Re Doesn't make any difference what OS.

          It does. For example, we are seeing some MIPS-based machines coming out of China, which look like they might be even cheaper than ARM, and at least as power-efficient.

          Linux is the only OS in common use that will run across x86, ARM and MIPS, and whatever else you might think of.
          ldo17
          • Try buying Oracle licenses some time

            Or most other enterprise level software. They typically charge per CPU so a server with many less powerful CPU's costs more to license than a server with a smaller number of more powerful CPU's

            Microsoft actually comes out very well in these comparisons since with their products they count each socket as a CPU. So when I install SQL Server on my Dell hardware with dual 6 core processors, Microsoft counts that as a two cpu installation even though there are 12 total cores.
            cornpie
          • Re: Try buying Oracle licenses some time

            Neither Oracle nor Microsoft make stuff that runs on ARM or MIPS, so I'm not sure what kind of point you're trying to make here.
            ldo17
          • So, what will Windows 8 RT run on? It won't be Intel, and it won't be MIPs

            so, what does that leave?

            In case you haven't heard, Windows 8 Rt is designed for the ARM architecture. Look it up.
            adornoe
  • Not hardly...

    "These workloads place the emphasis on light computing tasks but with lots of parallelism and frequent use of large amounts of data."

    Umm, no. As more and more applications are moved to the net and a browser, the back end processing power goes up, not down.
    NoAxToGrind
  • Elephant in the room...

    Since we are talking about what servers are going to look like a fet years out, it might be useful to consider what the coming fast-persistent storage technologies (spin, PCM, etc.) will do to CPU architectures, power consumption, physical size and performance.

    The effect of reducing the latency across your memory hierarchy is to reduce the value of CPU features that are there to maximize performance where latency increases with storage size. Cache, CISCO instruction sets, speculative execution and similar features are, at the end of the day built into CPUs to increase performance in the face of latency that increases as the data / code size grows. To the extent that fast-persistent storage reduces this latency, the value of CPU features designed to defeat latency is diminished.

    The real, practical question may not be how well or poorly an ARM processor can build an x86 like server, but rather which architecture will have the advantage when coupled to the coming fast-persistent storage technologies.

    Skate to where the puck will be..
    z2217
  • Good luck writing off Intel's dominance

    in the foreseeable future. They're not going to give up sacred ground without a fight. Moreover, their flex reserves run deep.

    Their name alone amounts to something of a standard.
    klumper
  • pulling the strings from either side...

    the fact that TDP is low, comes from the very fact that it packs less punch. But that will be surpassed in case of massively parallel designs and workloads. Packing more stuff and 'bits' and pieces to the ARM core means increased power consumption and a move away from the massively parallel scheme and towards much more of a scale up path. The designs around ARM, hardware and especially software for 'extreme scale out' OR parallel'ism' is the key to its success.
    sreesiv
  • unfair and unbalanced

    so, you interview Simon Segars, ARM's general manager of its processor and physical IP divisions, and suddenly know what the industry is doing???

    yikes, did you even call any of the many places that sell servers and compare what the product volumes are????
    CaptOska
    • Finally some one pointed that out...

      I was wondering if anyone else noticed it.
      TheBottomLineIsAllThatMatters
    • Article gives ARM's position, not industry's

      Hey CaptOska
      I wrote this article to delve into ARM's viewpoint on what the industry is doing. I talk with chip companies all the time as part of my beat and feel that ARM's point of view is legitimate. Nowhere in this article do I put chip volumes because as it stands at the moment ARM on servers is totally insignificant, but I believe in 2013 when we get 64-bit chips a change will occur.

      ARM acknowledges this weakness as well - ""I think you can regard ARM in the cloud as a player in 2014,"

      Furthermore, I went and had a chat with Facebook at their HQ on Friday. They are running large-scale tests in their datacentres with ARM and other non-x86 chips.
      Jack Clark