Intel Atom CPU anchors ultra-dense HP server

A standard 4U rackmount enclosure measures 7" tall, 19" wide, and up to 24" deep. How many servers do you think you could cram in one of those? 20? 100? How about 288? That’s the maximum density of Gemini, a high-density server spawned by HP’s Project Moonshot initiative.

Each one of the "server cartridges" contains a server-optimized Intel Atom processor code-named Centerton. The 64-bit chip has hardware virtualization support and takes advantage of ECC memory, making it well-suited to enterprise environments. Although Centerton isn’t listed on Intel’s site, the new Atom was discussed at IDF Beijing last month. It reportedly has dual cores and a 6W thermal envelope. 32-nm fabrication technology is used to build the chip.

Project Moonshot aims to reduce server complexity, which seems contrary to packing more processors into a rack. HP’s approach here is pretty straightforward: "move from tens of servers per rack that share basically nothing to thousands of servers per rack that share almost everything." Each Gemini cartridge features an Atom SoC, independent memory, and either direct or network-attached storage—that’s it. The chassis’ cooling, power, networking, storage, and management resources are all shared by the cartridges within.

According to HP, there’s considerable potential for energy, cost, and space savings. Even cabling is reduced. I’d expect the server cartridges to be easy to replace, as well.

If this sort of high-density server tech sounds familiar, that’s because it’s similar to the interconnect fabric developed by SeaMicro. AMD bought SeaMicro back in February, and at the time, it promised Opteron-based solutions in the second half of this year. The next logical step is micro-servers based on processors derived from AMD’s low-power Brazos platform.

Comments closed
    • DavidC1
    • 7 years ago

    It’s 288 compute nodes. I think that’s same as the ARM-variant.

    They are claiming some fantastic numbers, but its basically a corner case scenario. HP is saying that it will only take 12-14W while similar regular Xeon server will take 150W.

    They didn’t mention that the Xeon server would be able to handle much more workloads though, and that 150W is likely with the CPU idling most of the time, with rest of the platform using power instead. I bet if the same “regular” Xeons were fit with the optimized memory/storage/network these devices are using, that 150W will plummet too.

    I think it makes sense if you are a very small business and don’t need the extra capability for web serving(the exact scenario they are proposing this for use). But for others, there’s no advantage against regular core Xeons, especially those new Ivy Bridge ones with 17W TDP.

    • puppetworx
    • 7 years ago

    This could potentially reduce overhead massively given how little energy these chips use and how little heat they create. A server based on these could virtualize a lot of enterprise terminals (do they still do that?). More interesting is that combining a lot of these cheap, efficient, yet weak chips could provide the same computing power at a lower upfront and ongoing operational cost.

      • JustAnEngineer
      • 7 years ago

      Unfortunately, they’re probably becoming more common rather than less. My estimate is that the slowness, low display resolution and program incompatibility of the horrid Atom-based thin clients is costing our medium-sized company between one and ten million dollars per year in lost productivity compared to $100K worth of hardware and support cost savings.

    • fishyuk
    • 7 years ago

    We are doing ARM too don’t forget. HP is all about the right horse for the right course 🙂

      • Deanjo
      • 7 years ago

      [quote<] HP is all about the right horse for the right course :-)[/quote<] That's what Mister Ed said.

      • willmore
      • 7 years ago

      That’s not what you guys told us when we asked where the next gen of Alpha servers was. Back then it was “you’ll use what we offer or else”. Fortunately our management was smart enough (and our company big enough) to say “I think we’ll take ‘else’. Goodbye.”

    • ew
    • 7 years ago

    Is that 288 CPUs or 288 cores?

      • cynan
      • 7 years ago

      CPUs

        • ew
        • 7 years ago

        Are you sure? It says they are 6W each. So 288 of them would be 1728W and that is just for the CPUs. Would seem a lot more reasonable if it were 144 dual-core CPUs for 864W.

    • Bensam123
    • 7 years ago

    Makes you wonder what would happen if this was made with ARM chips instead of Atoms… Well besides a lot of software simply not being compatible with it.

      • ew
      • 7 years ago

      [url<]http://content.dell.com/us/en/enterprise/d/campaigns/project-copper[/url<]

    • Duck
    • 7 years ago

    Who would want one of these? Webserver hosts?

      • Bensam123
      • 7 years ago

      I’m guessing supercomputing. These wouldn’t be used for normal server applications as a lot of things aren’t neatly threaded into tens of thousands of threads.

        • OneArmedScissor
        • 7 years ago

        While I neither build nor program super computers, I think 300ish of these would have trouble keeping up with a few Tesla cards using the same amount of power.

        But there are certainly a lot of businesses who manage their own database, with employees connecting only intermittently, looking up one tiny amount of data at a time. The rest of the time, they’re just sitting around, and they certainly won’t need to be replaced with something even more powerful yet.

        There may not be much selling point for an “upgrade” to someone in that position, so anything new would have to actually save money overall to be worth replacing. That’s where something as dirt cheap as Atom comes in.

        After an update or two, maybe Atom [i<]could[/i<] be appropriate for web hosting services in the future. However, since those companies already put as many sites as possible onto as few servers as possible, they're going to be a lot more active and dependent on power efficiency.

          • Duck
          • 7 years ago

          [quote<]After an update or two, maybe Atom could be appropriate for web hosting services in the future. However, since those companies already put as many sites as possible onto as few servers as possible, they're going to be a lot more active and dependent on power efficiency.[/quote<] I was thinking with atom, each website could have it's own CPU. Maybe there is some advantage to doing it this way, such as it's easier to setup or manage than using virtualization to run many servers on much more powerful hardware??

        • Goty
        • 7 years ago

        Super computers are better served with solutions that are more efficient with regards to power required to complete a task (i.e. anything other than Atom).

        • UberGerbil
        • 7 years ago

        Only if it was extremely cheap relative to the competition. The fp performance would be so weak I can’t imagine any supercomputing project choosing it if they could afford one of the alternatives.

        Duck is right on the money with web servers. That’s an extremely common “normal server application” that makes productive use of thousands of threads. Apache, [url=http://www.zerigo.com/article/apache_multi-threaded_vs_multi-process_pre-forked<]for example[/url<], uses one thread per visitor (either all in a single process or spread across several processes). Say you're a web host with a few hundred active clients (sites) each with a dozen to a few hundred active visitors. Hundred x hundred = tens of thousands of threads. And you need extra capacity for sudden bursts of demand. So you need lots of processors, but their actual computing performance requirements are quite modest. And you can't charge a lot, yet need to have high uptime, so cheap density and low power consumption are important.

        • kukreknecmi
        • 7 years ago

        Around 100MFLOP of Atom core x 288 = 28GFLOP at best, which is the way lower than an i7 or i5. A Chinese 16 core ShenWei-W3, which is specifically designed for Supercomputing does 140 GFLOP at 1.1ghz(Averaging 8.75 GFLOP per core). So supercomputing / hpc is out of question.

    • moop2000
    • 7 years ago

    7″ Tall? That’s HUGE! 🙂

      • demani
      • 7 years ago

      <chuckle>Heh. Boys. </chuckle>

    • OneArmedScissor
    • 7 years ago

    I could have sworn they said these would be quad-core. I imagine some things might benefit from the massive amount of memory bandwidth, but it just seems kind of odd to split them across so many sockets when the goal is minimal power.

    That is, unless this is just the SoC for phones, recycled to keep costs down to the absolute bare minimum. That’s been a recurring theme for “new” Atom chips. :p

    I’d be very interested to see the dual-core SoC in tablets and laptops, though.

      • BobbinThreadbare
      • 7 years ago

      Could be dual core + hyperthreading.

        • OneArmedScissor
        • 7 years ago

        I’m sure they left the HT. For once, it makes some sense in this application. However, it turns out I was thinking of this likely nonsense about a possible 22nm Atom in 2013:

        [url<]http://semiaccurate.com/2011/11/01/intel-8-core-server-atom-gets-a-name-and-date/[/url<]

      • chuckula
      • 7 years ago

      The successor to these guys in 2013 will be a 22nm quad core part. Should probably slot into the same backplane for easy upgrades.

    • StashTheVampede
    • 7 years ago

    Would love to build an Atom based NAS using the latest Atoms. Heck, gimme dual NIC and one PCIe slot and I’ll make a router too!

      • Goty
      • 7 years ago

      Why not grab a cheap ARM board or a VIA Epia board and do the same thing in a smaller power envelope?

        • Deanjo
        • 7 years ago

        Because VIA Epia board are usually more expensive and harder to come by (plus have the disadvantage of using a via chipset) and the ARM boards usually don’t allow that much connectivity.

          • BobbinThreadbare
          • 7 years ago

          Don’t forget you lose out on all the x86 apps you might choose to run one day.

            • Deanjo
            • 7 years ago

            That would be least of a concern as the above mentioned uses are usually done on a headless system and using linux so there isn’t the same “OMG! It’s not x86!” limitations that you run into with Windows.

            • BobbinThreadbare
            • 7 years ago

            There are still plenty of limitations. Just because you have linux doesn’t mean you have every tool made for linux.

            • ew
            • 7 years ago

            Let’s see 28,125 installable packages in the amd64 version of Debian and only 27,556 installable packages in the armel version. OMG, you’re right there are plenty of limitations.

            Source: [url<]http://edos.debian.net/weather/[/url<]

      • Deanjo
      • 7 years ago

      I’ll plus one ya for the Atom based NAS, dual NIC and one PCI-e X4 slot.

      • Bensam123
      • 7 years ago

      Did somebody say Raspberry Pi?

      • kc77
      • 7 years ago

      I would still pick Brazos.

        • NeelyCam
        • 7 years ago

        Because a NAS really needs mediocre graphics?

          • kc77
          • 7 years ago

          No because even the latest Atoms struggle against Brazos. Atom sucks for CPU too. Only a fanboi more intent on supporting his/her on own bias would say something as stupid as “because a NAS really needs mediocre graphics” when obviously I’m not talking about that. Even so Atom sucks with that too.

          I’m a creature of benchmarks not of logos. So far all of the benches I’ve seen is of D2500 and D2700 struggling against E-350 much less the E-450 which easily surpasses the D2800.

          If Atom was really all that you would easily see the awesomeness of Atom broadcasted by everyone. It’s not awesome, which is why you don’t see tons of benchmarks on it.

          Don’t you think Intel would put marketing dollars behind a product which is awesome? Of course they would. They didn’t this time…. for a reason.

    • UberGerbil
    • 7 years ago

    This is what several start-ups like Smooth-Stone were attempting to do with ARM, but after a burst of publicity (and venture funding) a couple of years ago I haven’t heard much out of any of them. (Presumably they’re gated by ARM64, but they have some other hurdles too without the benefit of existing x86 server infrastructure)

    • demani
    • 7 years ago

    [i<]stoopid double post[/i<]

    • demani
    • 7 years ago

    This is interesting-I wonder what all those processors can do in as render nodes. I know there are some mighty burly blade systems, but the simplicity and CPU density of these might make it a worthwhile solution-576 cores should be able to make a nice dent at relatively low power (maybe 1.5Kw for a loaded chassis?).

      • willmore
      • 7 years ago

      Unless the core of these chips is drastically different from any Atom we’ve seen before, the performance/watt on floating point workloads will be poor. You’re better off with something that has few/faster cores for that workload.

    • dpaus
    • 7 years ago

    [quote<]I'd expect the server cartridges to be easy to replace, as well[/quote<] Assuming you mean 'upgrade' in addition to 'repair', hmmmm, veeerrrrryy interesting.

Pin It on Pinterest

Share This