Could Atom-based clusters put a dent in the server market?

You may be skeptical upon reading the headline above, but have a look at this story at VentureBeat about a start-up named SeaMicro. The firm has developed a server product that takes up one quarter of a traditional data-center rack and uses 512 Intel Atom chips for processing. By packing in power-efficient CPUs, SeaMicro aims to match Xeon- and Opteron-based servers while using less power and occupying less space.

The key to making it all work is SeaMicro’s special sauce, which reportedly is a combination of virtualization software and a series of custom chips that provide the customary I/O functions—storage, networking, managemenet—but do so through virutalized I/O that allows extensive sharing of hardware resources. How extensive? VentureBeat says, "Full told, SeaMicro eliminates 90 percent of the components from a system board. . . . With it, SeaMicro shrinks the size of the system board from a pizza box to the size of a credit card."

The SeaMicro SM10000 server. Source: MarketWire.

The article includes some power use and space estimates for a SeaMicro solution with a given SPECint_rate score alongside estimates for Dell R610 servers with similar performance. The bottom line is a claimed savings of over a million dollars during a four-year period. That’s a large installation, but if the math is sound, it proves a point worthy of note. For the right kind of applications, such as virtual web hosting, something like this might make a lot of sense.

Comments closed
    • NeelyCam
    • 11 years ago

    I kind of want to see what SeaMicro could do with 512 Moorestowns…

    • Crayon Shin Chan
    • 11 years ago

    Where was all this publicity when SiCortex was around? Now they’re out of business.

    • indeego
    • 11 years ago

    “SeaMicro claims that most of its customers aren’t bothered by the lack of ECC. There’s no hope for future ECC support unless Intel eventually embraces the Atom platform for servers.”

    Crazy. Especially on a server this large/importantg{<.<}g

      • d0g_p00p
      • 11 years ago

      First thing that came to my minds as well. Going to be a hard sell without ECC support.

    • kamikaziechameleon
    • 11 years ago

    Good innovative approach to the server space.

    • SNM
    • 11 years ago

    The storage on these machines needs to be completely disposable, which means you need NAS or SAS storage to be the primary store. I don’t know who this is targeted at. Amazon EC2?

      • jpostel
      • 11 years ago

      EC2 and other private/public cloud services are exactly right.

      Another good article: §[<http://gigaom.com/2010/06/13/seamicros-low-power-server-finally-launches/<]§ Server virtualization is all about increasing utilization of the server hardware to make more efficient use, so this system is right in line with that goal. Hardware with a lower power usage at idle is also gaining more visibility. Just like the TR folks show when comparing CPUs, you have to look at not only how long it takes to complete a task, but also how much power is required to complete the task.

    • Corrado
    • 11 years ago

    For the minimal price difference, why not use Core2 based CPUs and have FAR superior performance?

      • Fighterpilot
      • 11 years ago

      l[

      • Bombadil
      • 11 years ago

      I don’t think any of the Core 2s support the low power bus used by the US15 chipset (2.3 W TDP). This really is more power efficient than anything could be based on Core 2s.

        • OneArmedScissor
        • 11 years ago

        The Core 2 CULV platform could be rigged up to be more power efficient than Atom in pretty much any given scenario, but good luck fitting a few hundred instances of it in one box. :p

      • OneArmedScissor
      • 11 years ago

      Atom chips are tiny, have their own memory controller, and make hardly any heat.

      You are comparing apples and oranges. “Superior performance” has literally nothing at all to do with speed in this case.

        • Kurotetsu
        • 11 years ago

        A few of the commentators bring up the interesting possibility of an ARM-based server using the same structure (which should bring higher performance with, at least, equal or lower power to Atom).

        • bhtooefr
        • 11 years ago

        Not in this one they don’t – the Z530 uses a traditional chipset, it’s just that the chipset in this case is a single-chip north/south bridge.

    • LaChupacabra
    • 11 years ago

    I’ll file this under the “I’ll believe it when I see it” category of hardware.

        • LaChupacabra
        • 11 years ago

        Touche, how about “I’ll believe it when an unbiased third party (or me, personally, which is WAY less likely to happen) reviews the unit and is able to post benchmarks and tests proving that their are situations or workloads that would run more efficently at the same performance on this system.”

          • OneArmedScissor
          • 11 years ago

          You seriously need proof that Atom could conserve power compared to Intel’s multi-socket server platforms, which have all idled at several hundred watts for the past few generations?

          While a scientific comparison would be nice, running benchmarks isn’t going to give you one for something that’s meant to show its advantage when it’s *[

            • LaChupacabra
            • 11 years ago

            I’m sure an atom uses less power. Just like a 1.6 liter Honda engine uses less gas to go 200 miles than, say, a 560 HP Detroit Diesel DD15. But if the workload is to move 80,000 pounds of timber that same 200 miles, the diesel would take way less trips, and therefore use less gas. So, and let me quote:

            “tests proving that their are situations or workloads that would run more efficiently at the same performance on this system.”

            If it takes ten times the time to run something, and only giving you four times power savings, you’re still spending more on energy over time.

            • PrecambrianRabbit
            • 11 years ago

            The linked site does show better energy/performance than other server systems: 72 KWH to achieve a SPECint_rate score of 100,000, whereas a Dell R610 requires 254 KWH. I have no idea if the R610 is a good server by this metric, since I don’t work in the server space at all, but at least it improves over something.

            That said, using SPECint_rate is a little dubious. SPEC_rate is basically a test of how many completely independent, single-threaded workloads a system can run in a given time period. The single-threaded workloads in question are designed to stress the CPU and memory subsystem, rather than storage and interconnect. It’s basically the ideal case for tons of energy-efficient processors. A more realistic workload, however, may bog down due to other bottlenecks in the system. Plus, while throughput may be high, latency could be unacceptable.

            So, there are certainly lots of unknowns with this system, but I’m really excited about its existence… I’ll be curious to see how it does in the market.

            • bthylafh
            • 11 years ago

            I’d be more interested in how this thing does for I/O, especially with so much hardware virtualized away.

        • mako
        • 11 years ago

        The article at Anandtech is pretty interesting. There’s a nice picture of what evidently is a 32x PCIe card.

Pin It on Pinterest

Share This