Intel reportedly readying next-gen Atom with eight CPU cores

This past summer, we told you about Project Moonshot, a high-density 4U server packed with 288 Atom processors. Code-named Centerton, those Atom SoCs feature dual cores, single-channel memory controllers, and PCIe connectivity. They lack integrated Ethernet, Serial ATA, and USB controllers, though. Those  omissions will reportedly be addressed by Avoton, a new Atom SoC that CPU World says will incorporate logic to feed six SATA ports (four 3Gbps and two 6Gbps), four USB 2.0 ports, and four Gigabit Ethernet connections.

There will be more to Avoton than just extra connectivity. The chip will purportedly include up to eight CPU cores based on the upcoming Silvermont architecture, which is expected to bring out-of-order execution to the Atom for the first time. Those cores will apparently come in pairs that share 1MB of L2 cache. According to CPU World, clock speeds could reach as high as 2.4GHz with a 2.7GHz Turbo peak.

Rumor has it the memory controller is due for an upgrade, too. Avoton is supposedly slated to include a dual-channel controller with support for DDR3 and DDR3L memory at 1600MHz.

Avoton’s power envelope could climb as high as 20W, CPU World says, but that doesn’t seem entirely unreasonable for an eight-core chip with so much platform integration. The site claims Avoton will offer more performance per watt than Intel’s upcoming Haswell CPUs, which would be quite a feat. I imagine that applies only to very specific workloads—probably ones that are important to folks building high-density servers. We’ll know more in the second half of next year, which is when Avoton chips are supposed to arrive.

Comments closed
    • Diplomacy42
    • 7 years ago

    I’m starting to think Intel R&D has staged a coup and that they are just going “geeks gone wild.”

    • DavidC1
    • 7 years ago

    “The site claims Avoton will offer more performance per watt than Intel’s upcoming Haswell CPUs, which would be quite a feat.”

    How is that a “feat”? Regular Ultrabook bound Haswell has a TDP of 15W, and Tablet optimized versions are at 8-10W. The Clover Trail Atom coming nearly a year earlier has a TDP of 1.7W. That means for Haswell to have equal perf/watt, it needs to outperform Atom by anywhere from low as 5 times, to 9 times.

    Both are dual cores, Atom isn’t that slow.

    The more important thing is that CPU World also claims much higher performance/watt than the upcoming Centerton Atom chips. Centerton is the code-name for the new micro-server optimized Atom, rumored to be branded “Atom S”. Centerton is likely building upon Intel’s ultra low power Medfield foundations, and adds cool things like 8 PCI Express 2.0 slots for expansion chips(even though its a full SoC and can operate without a chipset), Virtualization support, ECC support, Memory RAS, and 64-bit. Basically its Medfield for servers.

    Now improving significantly over Centerton I can live with.

      • Anonymous Coward
      • 7 years ago

      [quote<]How is that a "feat"?[/quote<] I've never seen anyone demonstrate an Atom that can do work more efficiently ("task energy") than a proper desktop CPU which is running at an efficient speed. In fact I haven't been impressed by the idle power of Atom or Bobcat either, when compared to a properly optimized system containing a normal sized desktop CPU. So if a new flavor of Atom can actually do more work per watt than Intel's latest and greatest, I've call it a feat.

        • DavidC1
        • 7 years ago

        That has nothing to do with efficiency of a CPU core. Clover Trail for example, will have super low idle power for the entire platform. It’ll be easier if I say “ARM-class”. Then the TDP differences will start to matter.

        Nothing changes the fact that even the lowest TDP Core CPUs have 7x the TDP of average Atom CPUs. Which will matter when you do something on load, and very important when platform based power management is stellar.

    • BaronMatrix
    • 7 years ago

    Does this mean 8X as much waiting…? Or just 4X…?

    • Squeazle
    • 7 years ago

    Too many cores that come in pairs? Taking a page out of AMD’s book then. Hopefully they’ll pull it off a bit better.

      • internetsandman
      • 7 years ago

      All they share is the cache, it’s not like AMD’s bulldozer architecture where a significant part of the traditional CPU core is shared between two cores. Cache sharing between cores isn’t a new practice

    • aim18
    • 7 years ago
    • Bensam123
    • 7 years ago

    What is the clock/clock performance when comparing it to a normal i7 or whatever else you would want to cram into a server? I don’t know why they would want to start throwing massive cores at a atom, which traditionally doesn’t perform all that well and is just used in efficiency scenarios.

      • dmjifn
      • 7 years ago

      I’m interested in hearing about this too. The TR power usage charts show those i7s as being great for work per joule (when run flat-out), right? So if Atom is going to be awesome, is it because power consumption dropped more than processing power? Or are they basing this on total uptime, including idle and light workloads?

      That rumor about Amazon buying TI’s OMAP division may be bunk or it may be just for their Kindles. But if Atom is fit for the server space, maybe OMAP is not much of a stretch here either.

      • curtisb
      • 7 years ago

      [quote<]atom, which traditionally doesn't perform all that well[/quote<] This is largely due to in-order execution and a single memory channel. If the rumor is true and they are indeed changing to OOE and a dual-channel memory controller, the performance [i<]should[/i<] increase quite a bit for Atom. Here's to hoping it' works out that way and is a top (servers) to bottom (smartphones) change for the architecture.

      • UberGerbil
      • 7 years ago

      The proper comparison for this market isn’t clock-per-clock — it’s pages served or client response latency, both in absolute terms and per dollar or per watt (or joule). When you get up into the 8 core per package range, your’re targeting throughput servers not HPC, so the pure single-threaded performance (and, implicitly, raw clockspeed) isn’t as important.

      It was already clear from the emphasis Intel has shown with Haswell that it considers ARM a threat, and this looks to be just an extension of that. In the same way that Haswell appears to be particularly aimed at keeping ARM out of “ultrabooks,” this looks to be aimed at keeping ARM ouf of the server room, staving off the likes of Calexda and Smoothstone.

    • I.S.T.
    • 7 years ago

    This is a server CPU, and people are talking about this like it’s something we’re gonna be sold. Jeeze…

    As for the person wondering what’s the difference between this and Ivy: Well, devil’s in the details. OoO and all those terms you rattled off are high level descriptors, they don’t go into minute detail of how a processor works.

      • Ringofett
      • 7 years ago

      It’s possible we’ll be able to buy it; server-grade Xeon and Opterons are available, though I think 90% of the retail sales like that go to Folding@Home participants with bottomless pockets..

    • jjj
    • 7 years ago

    Avoton is a rather unfortunate name.

      • fantastic
      • 7 years ago

      Agreed. I say cut the marketing people down to one person and use common names.

      Replace “Core i” or “Phenom II X4” with Blue, Steve, Belinda, Wanda or Derek or something common.
      “Man, that new Intel Blue 5 looks like a nice chip.” The brand is the company name and the product name isn’t nearly as important.

    • Beelzebubba9
    • 7 years ago

    Maybe Intel should ask AMD how packing 8 slow cores onto a single die worked out. 🙂

      • bcronce
      • 7 years ago

      Maybe Intel should ask ATI how packing 1536 cores on a single GPU worked out. 😉

      It all comes down to design. The idea of lots of cores is fine if you know how to implement correctly and use it correctly, while understanding the trade offs.

        • bcronce
        • 7 years ago

        TR has so many post griefers. If you can beat them…

        • Beelzebubba9
        • 7 years ago

        Keep in mind Intel already has their own massively wide vector processor – Knight’s Corner – to compete with AMD and nVIdia’s GPUs, and and 8-core Atom certainly isn’t it.

        I (obviously) can’t speak for all sectors of the market, but virtualization really puts a dent in the many-core argument for most software and applications. Using a software hypervisor to dynamically allocate clock cycles for many servers or applications running on a smaller number of much faster cores means that you can have the dense provisioning for many concurrent applications you’d get from a many-core system, but also the option of allocating a substantial amount of execution power to a given task or server should that need arise.

        Obviously there are some niche applications for such a chip or Intel wouldn’t be building it, but for the types of workloads I know well, it makes a lot more sense to use fast 2S Xeon boxes and carve up your hardware at the hypervisor level.

      • demani
      • 7 years ago

      Don’t they already have their own to refer to to see how to do it right? Slow is relative, and at 20W there is room for a whole lot of cores to meet one Xeon or any other 8 core chip. So for servers and such it is a great opportunity (and hopefully also be cheap-there are a lot of multithreaded applications that would be great for this: storage would be one). Of course Avoton is going to need a lot more RAM to utilize it, so maybe intel will stop limiting that.

        • Beelzebubba9
        • 7 years ago

        The problem is this chip will be competing with Intel’s own Haswell based CPUs, and they’ll provide a massive uptick in performance:watt.

        As I said in the post above, a modern virtualization implementation removes a lot of the advantages of a many-core CPU since you can dynamically allocate the hardware resources available to any given system or application based on need.

          • OneArmedScissor
          • 7 years ago

          Haswell may not be in Xeon form until nearly 2015. It’s also extra high TDP and not as tightly integrated.

          That could be the most decisive schism between server parts that Intel has ever had. The “E” version was for niche applications that could justify the cost before, but now, maybe the same business will find a use for both platforms?

          Performance per watt probably isn’t the right metric for an Atom SoC. Maybe “storage devices per watt” or even “idle time per watt” would be more appropriate? :p

            • Beelzebubba9
            • 7 years ago

            [quote=”OneArmedScissor”<] Performance per watt probably isn't the right metric for an Atom SoC. Maybe "storage devices per watt" or even "idle time per watt" would be more appropriate? :p [/quote<] This is a good point. I know in our datacenters, CPUs are pretty low on the list of power drivers (a single SAN can suck down more power than a rack of servers), but I can only speak for what I know.

      • TaBoVilla
      • 7 years ago

      this comment is pure genious, why would people downvote this.. =(

    • Chrispy_
    • 7 years ago

    Bulldozer had 8 ‘cores’ but that didn’t stop it from being a significant step [i<]backwards[/i<] in most situations. The real question is IPC performance; Eight cores will be a disaster if those cores aren't strong enough to run single-threaded tasks in a timely manner.

      • Majiir Paktu
      • 7 years ago

      Like Geoff said: different workloads. Server farms aren’t trying to run mostly-single-threaded games at max settings. Single-threaded performance is important, but not in the same way.

    • nafhan
    • 7 years ago

    Interesting. I wonder if Intel’s long term plan will be to pull Atom and Core together, keeping that future CPU closer to the design of Atom, similar to what they did with Core and P4?

      • Diplomacy42
      • 7 years ago

      it seems like there goal is to have 1 processor family that runs anything from the server space, down through the tablet space and into the “things that don’t even need a processor” space… like coke machines and toasters.

    • HisDivineOrder
    • 7 years ago

    8 cores is 8 cores? Or 8 cores is 4 cores with hyperthreading?

      • tfp
      • 7 years ago

      Intel doesn’t state hyperthreading is a full core.

        • nafhan
        • 7 years ago

        Intel is good about that, but sometimes enthusiastic “industry experts” will call it a full core, and I’m having trouble finding something from Intel stating the number of cores.

        8 real Atom cores with a 20 W power envelope seems reasonable, though. So, I’m inclined to believe.

      • Squeazle
      • 7 years ago

      8 cores in 4 groups, which will [hopefully] handle things more efficiently than hyperthreading.

    • zdw
    • 7 years ago

    ECC support?

    I’d love to see an ECC-capable Atom for NAS use. Memory is so cheap these days that throwing an extra parity chip in is almost free.

      • Flatland_Spider
      • 7 years ago

      I think I saw somewhere that said yes since super dense servers and appliance are Atom’s market now.

    • sschaem
    • 7 years ago

    Since density is Seamicro business, it would make sense for AMD to use Avoton.

    This could turn out to be a more profitable business for AMD… use Intel / ARM CPU vs designing its own.

      • Antimatter
      • 7 years ago

      Surely AMD could build their own low power 8 core chip using the Jaguar core. At 3.1 sq.mm on the 28nm process even 16 cores would easily be possible. The Register suggests that AMD will have server chips with the Jaguar core.

      [url<]http://www.theregister.co.uk/2012/08/29/amd_jaguar_core_design/[/url<]

        • sschaem
        • 7 years ago

        But at what cost? when? and will it really be better ?

        It doesn’t make sense to spend 1+ billion a year in R&D if the result is non competitive products ?
        Wouldn’t It be cheaper for AMD to use Intel processors… they also then get access to 14nm parts.

          • shank15217
          • 7 years ago

          And Brazos is faster than atom by far. They will double the density next year. They need to fight Intel in the new space now because they have a strong current product and a well defined path.

            • NeelyCam
            • 7 years ago

            [quote<]And Brazos is faster than atom by far.[/quote<] Link? Because what I remember seeing was that Brazos was just slightly ahead in perf/clock, and behind in perf/watt (I don't have a link either now - I'd have to go back to old Anandtech reviews or something.. I was hoping you'd have a link, though)

            • Ringofett
            • 7 years ago

            [url<]http://www.anandtech.com/bench/Product/328?vs=110[/url<] Not the latest silicon in either case, but relevant I think. Brazos looks like a clear winner, at least circa early 2011. No power consumption on that page, either.

            • NeelyCam
            • 7 years ago

            Yep, that’s what I was thinking: Brazos is slightly faster clock-for-clock.

            Power efficiency was more difficult to determine as Intel’s platforms at that time were idling at unreasonably high levels, but if estimating the CPU power draw under load as the difference between system idle and load power levels, Atom looked more efficient (and that was back then when Atom was at 45nm).

            • shank15217
            • 7 years ago

            brazos faster clock for clock and perf per watt, its tdp is measured with the gpu in mind. For computational loads its better. Atoms ht helps a lot with io heavy workloads. We are actively looking at both these platforms for clustered nas and the brazos CPU is better for the encrypted and compression nodes but the atom worked out faster for general file serves because it could handle more requests in flight due to 4 hardware threads.

            • NeelyCam
            • 7 years ago

            [quote<]brazos faster clock for clock and perf per watt[/quote<] Link? [quote<]We are actively looking at both these platforms[/quote<] I guess this is based on internal testing, then. Would you be able to share some numbers? Are you looking at 32nm Atoms or 45nm ones? Also, in your evaluation, did you take look at full system power (which system?) or just CPU power? (I would imagine full system power.) Why that may be important to this discussion is that if Intel is going to offer an Atom server platform for microserver market, I would guess they'll improve system idle power.

            • Beelzebubba9
            • 7 years ago

            Brazos performance:

            [url<]http://www.anandtech.com/show/4134/the-brazos-review-amds-e350-supplants-ion-for-miniitx/6[/url<] Brazos power: [url<]http://www.anandtech.com/show/4134/the-brazos-review-amds-e350-supplants-ion-for-miniitx/7[/url<] Granted, I'm pretty sure that's 45nm vs. 40nm, but I'd imagine we'd see a similar story at 32mn.

            • NeelyCam
            • 7 years ago

            It’s tricky to get a good comparison out of this, as they didn’t use the same benchmarks for power and performance measurements. Also, as the platform idle powers are so high, I’ll try to estimate CPU power usage by looking at the difference between load and idle. This is my best attempt to compare:

            Cinebench (R10 for performance, 11.5 for power):

            Brazos: 2251/(32.2W-24.2W)=281.4/W
            D510: 2024/(38.6W-35.2W)=595.3/W
            [b<]Advantage: Atom (+109%)[/b<] H.264 video decode (720p for performance, 1080p for power); Brazos: 14.1/(33.4W-24.2W)=1.53/W D510: 13.3/(40.5W-35.2W)=2.51/W [b<]Advantage: Atom (+64%)[/b<] I understand this comparison is nowhere near perfectly accurate, but from this information I can't figure out a better way to do it..

            • Beelzebubba9
            • 7 years ago

            I guess either one of us could have just looked harder and got the numbers that we actually were hoping for. 🙂

            [url<]https://techreport.com/review/20401/amd-e-350-fusion-apu-on-the-desktop/5[/url<] So if you look at total task energy, using brick PSUs, it looks like the E-350 is a whopping 13% more efficient than the D525.

            • NeelyCam
            • 7 years ago

            Same thing as in Anandtech – power consumption is dominated by the rest of the system power. That D525 system idle power is ridiculous – it’s the same as in my Clarkdale i5-670 rig! Atom can most certainly be idling at lower system power levels if the rest of the stuff is designed to be low power

            Again, we’re left with comparing load and idle power differences to try to estimate how much that Atom CPU actually consumes.. in this case, the difference for Atom is 4W and for Brazos it’s 9W.

    • anotherengineer
    • 7 years ago

    LOLZ

    8 core Atom, hopefully code name isn’t Oxygen, that would be lame

    sorry couldn’t help it

      • nafhan
      • 7 years ago

      It’s at least as good as most other chip naming conventions…

      • BIF
      • 7 years ago

      Oxygen could be the shared L2, while the two core-pairs could be Hydrogen.

      I could see the ads now; “Performance as smooth as water!”

      😉

    • ronch
    • 7 years ago

    So two cores will share a meg of L2 cache. Hmm, reminds me of Bulldozer. What’s it called, a module? 😀

      • colinstu12
      • 7 years ago

      except these are both real cores. where BD it’s two FPUs per core (and somehow that is “two” cores. it really isn’t).

        • sschaem
        • 7 years ago

        For that market it doesn’t matter much. Integer/branching is key.
        You can see that in BD benchmark. where a 4 module BD ‘spank’ a 4 core Sandy Bridge.

        [url<]http://images.anandtech.com/graphs/graph4955/41698.png[/url<] This is what AMD was after. Allocate more resource for branching/integer workload. If they didn't do this, they couldn't fit 4 module on the die...

          • UberGerbil
          • 7 years ago

          Yes, as clueful observers have noted since its introduction, Bulldozer looked far better-suited to server loads than typical desktop uses.

          • Waco
          • 7 years ago

          4 SB cores [i<]with hyperthreading[/i<] aren't much behind the higher-clocked 4 module BD...

        • Goty
        • 7 years ago

        It’s actually two IPUs and one FPU per module. You should probably know the basics if you want to be a successful troll.

        • ronch
        • 7 years ago

        And yet, I’m still sure those BD cores will outgun those Atom cores.

          • NeelyCam
          • 7 years ago

          Not at the power levels we’re talking about here

            • ronch
            • 7 years ago

            Obviously, but I kinda sensed we were about to embark on another ‘Bulldozer cores are not real cores’ discussion.

            • NeelyCam
            • 7 years ago

            They aren’t 😛

      • sschaem
      • 7 years ago

      I also believe AMD core architecture is not off…
      But so far their 16 core solution (2.3ghz) is rated at 115w (Its also a beast vs an Avoton)
      So 57 watt for 8 core. Even assuming BD get 20% higher IPC , AMD need to double power efficiency to be competitive with this new breed.

      I have no clue as to what AMD will get with a potential 28nm Piledriver opteron.
      Hopefully Seamicro Opetron based product will offset this CPU inefficiency delta

        • Anonymous Coward
        • 7 years ago

        I think even BD gets well over 20% higher IPC than an Atom.

      • Anonymous Coward
      • 7 years ago

      I would not be surprised if a shared FPU was a good idea for an Atom-style processor.

    • chuckula
    • 7 years ago

    [quote<] The site claims Avoton will offer more performance per watt than Intel's upcoming Haswell CPUs, which would be quite a feat.[/quote<] Considering that performance/watt is already in Intel's favor compared to ARM, if Avaton beats Haswell there will be strong competition waiting for ARM's first 64-bit chips.

      • NeelyCam
      • 7 years ago

      The original Atom offered more perf per watt than Nehalem. It’s easier to offer high efficiency with low-performance chips. For example, with lower clocks you can scale down supply and get an extra efficiency boost. It’s much harder to scale up performance without giving up a lot of efficiency.

    • Game_boy
    • 7 years ago

    If it has high clocks, out-of-order, dual-channel memory, presumably a ringbus and L3 to support 8 cores’ traffic, then how is this a different architecture to their desktop CPUs?

      • tfp
      • 7 years ago

      It doesn’t have high clocks, it is much lower power, and each core is not wide like Ivy.

      • UberGerbil
      • 7 years ago

      2.4Ghz is “high clocks”? Even if the high-level feature-set is similar (and it may not be as similar as you think when it comes to execution units), and even setting aside the differences due to far greater SoC integration (which you shouldn’t, though it appears Broadwell may approach that from the other direction) they are likely to differ in a far more fundamental way: the transistor libraries used. The Atom-based designs will likely use libraries and design strategies optimized to reduce leakage rather than maxmizing clockspeed headroom; that they’re contemplating something as high as even the mid-2GHz range probably says as much about the capabilities of their 22nm finFETs as about the actual design.

        • MadManOriginal
        • 7 years ago

        It’s also not really a *huge* increase over current Atoms which top out at 2.13 GHz.

Pin It on Pinterest

Share This