IDF keynote reveals new server processors, rack architecture

The year’s first Intel Developer Forum has kicked off in Beijing, China. Today’s keynote address is filled with details about Intel’s future server processors, including fresh additions to the Atom and Xeon families, plus a new design concept for the next generation of datacenter servers. Let’s start with the rack-scale architecture, which Intel sees as the future for the datacenter.

Today, server racks are typically filled with largely complete systems that may share power and cooling components but little else. Intel wants to break things up, separating the core processing platform from the storage and networking elements. Those base building blocks would be joined by an ultra-fast switch fabric based on silicon photonics technology. This disaggregation is expected to make servers more flexible, cheaper to operate, easier to manage, and more efficient overall.  

Source: Intel

Intel is working with Alibaba, Baidu, Tecent, and China Telecom on an initiative called Project Scorpio that puts some of those design goals into practice. HP is moving in the same direction, as well. Last year, we learned that the PC maker would offer high-density Project Moonshot systems using modular cartridges based on Atom CPUs. The first Moonshot machines are now shipping with the Atom S1200 processors Intel announced in December. New Moonshot servers based on Intel’s upcoming Silvermont Atom refresh are expected later this year, as well.

The next generation of Atom processors, Dubbed Avonton, will purportedly allow HP to pack four times the number of SoCs into each Moonshot server. The individual Avonton chips should be more potent than the Centerton silicon from the Atom S1200 family, and Ethernet connectivity will be integrated this time around. Avonton is being be fabbed on the same 22-nm process as current Ivy Bridge chips. According to the keynote presentation, the new server-grade Atom is sampling now and will be available in the second half of the year. It will be joined by Rangley, a tweaked version of the SoC aimed at networking gear like routers, switches, and security appliances.

Speaking of specialized Atom processors, Intel used the IDF keynote to reveal the Atom S12x9, which is available now. Based on the current Saltwell Atom CPU architecture, this SoC is designed for storage products and features 40 lanes of PCIe 2.0 connectivity—five times the number of lanes present in Atom S1200 chips. The S12x9 family employs hardware-based RAID acceleration to reduce the processing load for the CPU cores. It also supports Asynchronous DRAM Self-Refresh, which "can protect critical DRAM data in the event of a power interruption." 

Source: Intel

The IDF keynote also dropped some details on several upcoming Xeons, including a new E3 family based on the Haswell microarchitecture. The Xeon E3 series is due in mid-2013 and promises TDP ratings as low as 13W. These chips will be limited to single-socket systems and offer up to four cores and 8MB of L3 cache. You can expect 20 integrated PCIe 3.0 lanes and support for up to 32GB of memory. The desktop versions of Haswell will probably bear a striking resemblance to the E3 CPUs.

Higher up the Xeon line is the E5 family, which is getting a refresh in the form of Ivy Bridge-EP silicon. These processors will sport up to eight cores and be available for dual-socket systems starting in the third quarter. Also on the menu: 20MB of L3 cache, 80 PCIe 3.0 lanes, and support for up to 768GB of system RAM.

The big daddy in the Xeon camp is the E7 family. Based on Ivy Bridge-EX silicon, these monsters will boast up to 10 cores per processor and support systems with four or more sockets. Intel says eight-socket systems can address up to 12TB of RAM, which is more memory than most folks have storage. The E7 chips will also feature 30MB of L3 cache and 144 PCIe lanes. They’ll have 130W TDP ratings, as well. The other Xeon chips cover a range of thermal envelopes from 13 to 130W.

Comments closed
    • Usacomp2k3
    • 6 years ago

    Too bad they can’t include a Larrabee module as part of this. Would provide some competition to AMD & nVidia’s Cloud GPU’s / GPGPU solutions.

    • willmore
    • 6 years ago

    Project Scorpio? Really? Is it run by Hank Scorpio? Why do they need a nuclear reactor?

      • Concupiscence
      • 6 years ago

      I still think it’s funny as hell that Hank Scorpio is basically Larry Ellison.

        • willmore
        • 6 years ago

        Holy heck, but that’s an apt comparison!

    • chuckula
    • 6 years ago

    While not mentioned above in this TR article, apparently Intel also released more details about upcoming Phi models. They have a new higher-end model at about 1.2GHz in the same power envelope as their earlier high-end models, and apparently there’s a single Chinese supercomputer that is using 100,000 Phis, so it looks like Intel has been able to sell a few here and there (http://vr-zone.com/articles/intel-xeon-2013-update–a-bit-later-but-a-bit-better-too/19581.html)

    • smilingcrow
    • 6 years ago

    Any old iron, any old iron, any, any, any, old iron.
    ARM! Cor blimey guvnor, please sirs can we have some more orders as we need to keep our fabs a churning.
    God bless ya.

    • ronch
    • 6 years ago

    Don’t get me wrong, I’m perfectly happy with my FX-8350, but man, look at those Ivy Bridge-EP and IB-EX dies. I wonder just how much Intel would price the EP 8-core if it were to sell this in the desktop space.. Surely, way over the $190 that AMD charges for its top dog.

      • Archer
      • 6 years ago

      Dog is right.

        • Airmantharp
        • 6 years ago

        Don’t hate.

      • sircharles32
      • 6 years ago

      You do have a choice: Arm or Leg? Also, you only need one kidney to survive………

        • ronch
        • 6 years ago

        Fortunately, AMD is around to keep me happy while at the same time allowing me to keep all four limbs and two kidneys, plus more cash in my wallet.

    • esterhasz
    • 6 years ago

    I see these developments as mainly a reaction to Facebook’s open compute architecture efforts (http://www.opencompute.org/). Companies like FB are not well served by technology designed for banks, insurances, and airline reservation systems.

    They do not operate data centers but networks of data centers. That means that redundancy, uptime, etc. have completely different requirements. Just imagine the faces at HP and IBM when FB announced that they’d design their own hardware. That’s a big alarm bell right there!

    As server development is much more of a collaborative effort between chip designers and the big server companies, there is enormous pressure on Intel to deliver architectures that make it easier to adapt to changing requirements and more quickly to serve a diversifying market.

    While discussions on tech sites see the ARM vs. Intel race in the framework of a sports game (passion++, analysis–), the simple reason for the actual debate is that the current transformations in application patterns create market openings that are so glaringly big that a social networking company feels it’s not absurd to build hardware on their own (or have Quanta built it to spec). That’s complete market failure.

    I have no clue whether Intel or ARM or somebody else will fill these gaps, but it is clear that server technology has to diversify much more quickly to adapt to the general diversification trends in IT. The reason why anybody gives ARM a chance in this is that traditionally, smaller companies are quicker to adapt to changing markets and the ARM model allows for a lot of the experimentation that is required to actually know where the trends are going. Throw a 100 ARM companies against the wall and see what sticks.

      • NeelyCam
      • 6 years ago

      [quote<]Throw a 100 ARM companies against the wall and see what sticks.[/quote<] Yes; this will be interesting. A 100 ARM licensees designing various things in parallel with small budgets, hoping to score big against a juggernaut like Intel with comparatively infinite resources (and a significant capacity, process and - yes - cost advantage to boot) but also arrogance and decades of tradition slowing down response time. If one of these ARM licensees stumbles across a perfect formula, some big company like Facebook might really give them a chance. The big question is, how cost sensitive is Facebook on a small part of the hardware (like a CPU). If the answer is "not that sensitive", Intel is the way to go. The parts will be expensive because Intel likes margins, but if it's only a fraction of the total system cost but gives a major performance boost, Facebook might not care. It's probably more crucial that Facebook gets to 'customize' the part better for their specific need (features and performance/watt for their application), and it's unclear if Intel is willing to accommodate something like that. My guess is that they would be, since Facebook is such a huge potential customer, and I've seen some EETimes articles claiming that Intel would be happy to do that, but then there's that arrogance... So, we'll see how it goes. My money's on Intel, literally

        • esterhasz
        • 6 years ago

        [quote<]So, we'll see how it goes. My money's on Intel, literally[/quote<] Mine, too. As a small time investor I would never bet on a startup. And they've got a great dividend. But sure, I don't think at all that Intel is on the way out. But there's room now for niche players to make a real dent - question is whether they can grow to size before they're snatched up.

    • chuckula
    • 6 years ago

    I’m sorry, did those losers at Intel try to post slides of chips? All I see are HUGE failure, BIG Failure, Hasbeen Failure, and ARM roadkill failure!

      • BoBzeBuilder
      • 6 years ago

      ^ I like this guy.

      • MadManOriginal
      • 6 years ago

      You need some new material.

        • chuckula
        • 6 years ago

        It was late. I was also going right-to-left for some reason.

      • internetsandman
      • 6 years ago

      You’d think a site like TR would play host to classier individuals and repel the common Internet troll

        • Srsly_Bro
        • 6 years ago

        His mom didn’t give him much attention when he was little. Give him a break!

        • chuckula
        • 6 years ago

        Oh… there’s nothing common about my trolling.

      • ronch
      • 6 years ago

      Are you the real chuckula? Or did you kidnap him and took his TR username and password?

        • chuckula
        • 6 years ago

        Chuckula doesn’t live here anymore!
        (insert spinning head & demonic levitation here)

    • melissanxel032x
    • 6 years ago
    • sjl
    • 6 years ago

    This is sounding very familiar. Oh yes: blade computers. Buy this great big 10RU box, and slot in fifteen or twenty “blade” servers from the vendor for higher density computing. Great for reducing power and redundant components; not so great for reducing costs (because once you’ve bought the blade frame, you’re locked into that particular vendor for new blades. Best of luck if they decide to move to a new blade architecture …)

    There’s actually a remarkable degree of overlap between this sort of thinking and virtualisation; looking, for example, at the IBM lineup, it can be a difficult choice to make between, say, a bunch of blades, or a single monolithic server running PowerVM and a significant bunch of LPARs (which can be a combination of AIX, Linux, and IBM i, all on the same frame). All in the name of saving overall costs on hardware – it’ll be interesting to see what sort of mix the market springs for in the long haul.

      • slowriot
      • 6 years ago

      Great points. If you look at IBM’s own Flex System you can see them leveraging PowerVM and a blade-like nodes (computer, storage, network) that plug into a chassis with the management bits integrated. The nodes can be swapped as needed and the workload instances (varying layers of virtualization available, could be instances of Linux or AIX WPARs) can dynamically grow and shrink across the available hardware resources.

      I’m looking forward to getting more details on Intel’s approach.

        • sjl
        • 6 years ago

        In some ways, the “honking great big server with virtualisation” approach is more flexible than the “lots of small blades” approach, especially if the virtualisation layer (as is the case with PowerVM) is rock solid. (x86 virtualisation, I suspect, has a ways to go before it gets there – it’s good, but it isn’t as transparent as other platforms. Yet.)

        But it depends, very heavily, on expected scaling requirements, required load levels, and server costs – eg, buying a single p795 for an entire company is a very dangerous move (since if the server goes down, the company is SOL). Hot spares are a great idea, but having expensive hardware sit idle is generally frowned upon by accountants (which is why, very frequently, the test/development environment also serves double duty for DR.)

        Both approaches, I suspect, will be very strong in the market for considerable time to come.

        Disclosure: not really a virtualisation admin, but I have done some work with AIX and PowerVM in a previous job. Fascinating topic in a great many ways.

          • esterhasz
          • 6 years ago

          Yeah, good points! But to add to the complexity of server discussions: these technology questions are just part of the equation – and sometimes even not the decisive one. The fact that many software packages in the server world (e.g. most DBMS) are still priced on the number of cores they will run on, introduces a variable that can weigh heavily on any decision.

          Compute units as calculation unit are shaking that up a little (e.g. an Atom core would count as 0.5), but the pricing dynamics still introduce a logic into the server world that makes server room architecture quite a complicated exercise.

          And indeed, the concepts touted by Intel are not new, but it looks like an effort in standardization which could reduce some of the vendor/platform lock-in you talk about.

            • sjl
            • 6 years ago

            Yeah; per core licensing models are horrible and deserve to die a long, slow, tortuous death. Especially when the software in question (TSM – the backup product I wrangle professionally – springs to mind) doesn’t really have anything significant CPU-wise to worry about. Completely agree with you on just about everything you’ve said there; if this leads to a standard blade-type architecture that can be used by Dell, HP, Gateway, Acer, IBM, and similar, I’m all for it. I doubt that HP or IBM would join in on their Unix platforms (POWER, Itanic), but even if it’s only x86, it’ll be a good start.

            And come to think of it, per core licensing can be a disincentive for virtualisation – lots of companies insist on licensing on the physical CPUs, without regard to the virtual CPU count assigned (“but you could crank it up while we’re not looking!”)

          • shank15217
          • 6 years ago

          oh please, what metric do you have that shows that x86 virtualization isn’t up to par, another delusional big iron user. IBM is losing business fast to x86 and arm. Even the vmware drones have a better bet than powerVM. x86 vm technologies have a huge variety and today’s vms can have access to bare metal for all io needs as well as support nested modes just like lpars

            • sjl
            • 6 years ago

            A five minute search shows, for example:

            [url<]http://securitytracker.com/id/1028068[/url<] [url<]http://www.securitytracker.com/id/1028101[/url<] [url<]http://securitytracker.com/id/1027018[/url<] (which, if I read it rightly, lets a user on a guest VM gain access to the hypervisor layer.) I am yet to see a report of a security vulnerability in PowerVM. That's not something that can be said about x86 virtualisation platforms.

            • shank15217
            • 6 years ago

            I don’t think that’s a fair comparison. Its like Mac users claiming that their os is safer because there are fewer viruses for it. Smaller foot prints and less available hardware means that its not a primary focus for security researchers as well, keep that in mind.

    • NeelyCam
    • 6 years ago

    Even if PC is dead, servers are alive and well, and Intel is the incumbent to beat. ARM assault has failed so far; Intel got to microservers first, and Avaton will only increase their lead. High-end, AMD might be able to finally bring something to the table again, but Intel will likely hold on to at least 90% of the market.

    I expect the stockprice to skyrocket; today was a good sign of things to come. Good thing about the “PC is dead! (no actually it’s growing)” fiasco from last week; I was able to increase my position by 500sh on the cheap

      • sschaem
      • 6 years ago

      How can a company grow, when its already totally dominating the market its in ?!

      And ARM did not start the assault yet. Yes, you have people like calxeda jumping the gun on 32bit ARM, but thats not the onslaught we will see in late 2014.

      I’m not stating that Intel will loose its dominant position, but its market share will start to shrink.
      All that it took away from AMD, and more, will be taken back.

      I would say that Intel today is at its absolute peek in term of market share.
      (the x86 phone market look like a dead start)

      And you never know, AMD using TSMC 20nm might do some damage, and Intel will lose its near total x86 desktop & laptop dominance.

      Seem like the only direction from here is down. (But the stock seem undervalued by 30%+ so we might see another decade of flatline. The stock was $24.5 in 1997 after all, and 21.75 today….

      So I hope you got the stock for the dividend, and not to make a quick buck.

        • MadManOriginal
        • 6 years ago

        A company can grow in absolute terms when the product market is growing even when its already totally dominating that market. It can even lose market share percentage and still grow.

          • Airmantharp
          • 6 years ago

          Economics, how the eff do they work?!?

            • MadManOriginal
            • 6 years ago

            It’s not even economics, it’s simple multiplication. 50% of a pie that’s 2x as large is the same absolute amount as 100% of a pie that’s 1x.

          • sschaem
          • 6 years ago

          Its obvious that if you own 100% of a market and the market double, you double your sale.

          This is not what I question. What I dont see is any signs that the server growth will accelerate.

          And its naive to think that the introduction of 64bit ARM will increase Intel market share.

          So what is the catalyst for the accelerated growth that doesn’t exist today?
          and what make you think the onslaught of 64bit ARM server chips and offering will increase Intel market share ?

            • MadManOriginal
            • 6 years ago

            As a matter of fact, Intel’s data center division sales have been growing slowly but consistently for quite some time. Search ‘Intel data center sales’, it’s just a simple fact. (I pay attention to quarterly results summaries since I own Intel shares).

            I also didn’t say that Intel’s market share will grow, just that a company’s sales can grow even if market share shrinks. It just depends on which changes faster.

            • jihadjoe
            • 6 years ago

            We’re generating more and more data each day.

            Email services are becoming increasingly like google and are now offering near-unlimited storage instead of forcing you to delete once inbox is full. People generate pictures, video and music and upload to social media and public sharing sites. Also, there’s cctv’s and big bro watching all of us!

            Even as more people change from analogue based systems to digital, the resolution of digital data is also increasing fast. Cameras have gone from < 1MP to over 30MP, video from poststamp sized clips, to 1080P and soon 4K.

            The demand for servers will only grow, and won’t stop growing until we stop making more data, which isn’t happening anytime soon.

            • Kreshna Aryaguna Nurzaman
            • 6 years ago

            Um, you mean the demand for SAN storage, don’t you?

            • jihadjoe
            • 6 years ago

            That yes, but you also need lots of servers to do analysis stuff with the data. It’s not just plain storage.

            cctv footage, for instance is in some places automatically analyzed to alert police of suspicious stuff left in public places, automatically tag cars for parking and/or moving violations.

            How much slower would search engines (and ad servers) be if they had to index an ever increasing amount of data, or serve ever increasing numbers of users (thanks to tablets and smartphones) but were stuck with a fixed amount of processing power?

        • NeelyCam
        • 6 years ago

        [quote<]So I hope you got the stock for the dividend, and not to make a quick buck.[/quote<] Dividend helps, and "quick" depends on your time horizon. I'm thinking longer term; in 10 years I'll be stinking rich. Like [i<][b<]Deanjo[/b<][/i<] rich. And then my socialist bleeding heart will feel guilty, and I'll give it all away

          • sschaem
          • 6 years ago

          10 years ago Intel stock was at 32$ … you better sleep with an eye open 🙂

      • ronch
      • 6 years ago

      Perhaps AMD’s upcoming Steamroller will provide a significant performance in terms of IPC and clock speeds. I estimate it to be at most 30%, nothing more. However, given Intel’s wide process lead it will be hard for AMD to keep up especially in terms of power efficiency. The least AMD could do is go straight to 22nm or thereabouts if it wants to be on rough parity with Intel until Broadwell comes out, but their inability to do that, and instead settle for 28nm on their next shrink, means Intel’s lead will become bigger for the foreseeable future.

      So, question is, can’t AMD get GF or anybody else do 22nm for them this year?

        • BoilerGamer
        • 6 years ago

        TSMC’s 20nm won’t be ready until 2014, and their volume would be highly limited due to GPUs that need to be produced on that process then(Maxwell/Volcanic Island). Nobody else AMD could use have CPU/GPU grade 20nm/22nm process ready in 2014. GF’s next die shrink is 20nm but there is no sign that it is remotely ready for use on AMD CPU(AMD had issues with GF yields/quality before) use in 2014.

          • Airmantharp
          • 6 years ago

          Pretty sure TSMC’s next node will be hogged more by Qualcomm making Snapdragons than consumer GPUs :).

      • ronch
      • 6 years ago

      [quote<]Even if PC is dead[/quote<] PC IS [u<]NOT[/u<] DEAD!!!!

        • MadManOriginal
        • 6 years ago

        ZombieC

        …say it fast and with lazy lips.

      • Stonebender
      • 6 years ago

      Expect the stock price to skyrocket eh? You must not have been paying much attention over the last decade. It doesn’t matter how well Intel does, the stock price barely moves.

        • NeelyCam
        • 6 years ago

        The stock price hasn’t moved because of sentiment (recession, missed mobile, PC dead). If you look at P/E, it has dropped drastically to keep the stock price down even though revenue and profits grew significantly.

        P/E of a stock like Intel can be below 10 only for so long. Now that PC is getting resurrected by Haswell, tablet market is getting taken over by Bay Trail, phone market taken over by Merrifield, microserver market by Avaton… sentiment will change, and stock price will move.

          • Stonebender
          • 6 years ago

          I’ve been with the company for thirteen years, and would love to see the stock start acting like it did in the nineties… I just don’t see it happening. If record quarter after record quarter of revenue growth doesn’t get the stock moving, nothing will. I don’t expect the stock to get over $30. It more than likely will hover around $25.

            • MadManOriginal
            • 6 years ago

            Of course it on’t move like the 90s, it was called a bubble for a reason. Intel peaked at ~$29 recently after a run of record results and that’s not an unreasonable price just based on average historical P/E of the broad market.

Pin It on Pinterest

Share This