10 Comments

Summary:

Facebook has made waves by detailing its plans to use what an executive calls chips that have a cell-phone architecture in its future data centers. The social network plans to test such chips now and next year and will likely have them in production in 2014.

Frank Frankovsky Facebook
photo: Pinar Ozger
Frank Frankovsky Facebook

Frank Frankovsky, VP, Facebook<br />(c)2012 Pinar Ozger pinar@pinarozger.com

Facebook has made waves by detailing its plans to use what an executive calls “cell-phone chips” — or “wimpy cores” — in its future data centers. Frank Frankovsky, the VP of infrastructure at Facebook, told me the social network plans to test such chips now and throughout next year, with plans to have them in production in 2014.

“We’re testing and generally bullish on the category, and based on some of the early testing our useful work per watt per dollar will improve, although that varies by workload,” said Frankovsky, “although our Hip Hop [loads are] the most CPU intensive and that hasn’t been ruled out.” HipHop is the open-source code Facebook uses to speed up the PHP code underlying the entire site.

But the question Frankovsky can’t or won’t answer is which of these cell-phone chips Facebook might adopt: a question that is hugely important, given the size of Facebook’s infrastructure and the influence it can exert on other companies as a founding member of the Open Compute Project. So let’s just do a quick rundown of the possible winners in this particular data center cage match.

Frankovsky was very clear in talking to me about what he thought of as a wimpy core. It doesn’t have to actually be an ARM chip to have the cell-phone-style architecture he referred to earlier. Intel’s Atom is still in the running. However, he did say graphics processors aren’t something he’s considering, because they don’t make sense for his workloads. On the issue of whether 64-bit-compatible chips would be in store, Frankovsky hedged, refusing to outright commit to 64-bit, but he said the social network doesn’t “plan to adopt anything that’s not 64-bit.”

As for timing, he said the testing is ongoing with some adoption in the latter half of 2013, and “if all the stars align then [Facebook's adoption] will be a material impact to the market” by 2014. So which companies might see the impact of Facebook’s adoption of wimpy cores? Let’s run down the contenders.

Andrew Feldman of AMD, Barry Evans of Calexda, and Guido Appenzeller of Big Switch Networks Structure 2012

(L to R) Andrew Feldman of AMD, Barry Evans of Calexda, and Guido Appenzeller of Big Switch Networks<br />(c)2012 Pinar Ozger pinar@pinarozger.com

Calxeda. This Austin, Texas–based startup counts Frankovsky as an advisor of sorts, but its product — a system that combines several ARM-based cores and a proprietary networking chip so those cores can communicate — is just off the line. Plus, it will have to wait until 2013 or 2014 until its systems can support the 64-bit instruction set. This is within the Frankovsky timeline, and the workloads he has mentioned are ones where Calxeda is trying to establish tests and benchmarks.

Marvell/Dell. Marvell is also using ARM-based cores in its Armada line of chips, and Dell has picked up those chips to start testing a line of servers. Dell announced the line in May, and it is already a huge supplier of hardware for Facebook as well as a participant in the Open Compute Project that Facebook founded. Plus, Frankovsky is a former Dell employee from the DCS group, where Facebook bought a lot of its hardware.

Intel. Frankovsky was very clear that he regards Intel’s Atom core to be in the class of wimpy cores he is considering. And while many may scoff, Intel has done a good job reducing the power consumption of its x86 chips for the Atom line, unveiling the Centerton chips this month that will ship inside HP’s low-power servers (HP also has a deal to bring Calxeda’s systems to market). The new systems will be able to do as much work as a traditional 150-watt system in a 12-to-14 watt envelope. Plus Intel already makes a 64-bit Atom part designed for SeaMicro, a company building low-power microservers. Intel has an existing relationship with Facebook, the dominant x86 architecture and a 64-bit part in the market.

AMD. This is a bit difficult to assess, given that prior to buying SeaMicro in March, AMD didn’t really have much of a story or option for wimpy cores. It still doesn’t, but buying SeaMicro, which uses Intel’s Atom part, gives it an entrée into the market that it will press. Plus, when I listed the competition Frankovsky told me, “I wouldn’t leave AMD out of this race either.”

Tilera. This Cambridge, Mass.–based startup has been building massively multicore chips designed for big data and cloud workloads since 2004. It has an advantage of having more than 10,000 existing cores running in production at unspecified customers, according to Ihab Bishara, the director of server solutions for Tilera. It has also been tried by Facebook in its 32-bit iteration and won favorable comments from the social network. Cynics have claimed Facebook did that test just to keep Intel on its toes, but Tilera has the first non-x86, 64-bit-based server available in the market and is already deployed in servers in 3 of the top 20 websites.

  1. How about something like Sun’s SPARC4 architecture, or even SPARC5 due out next year? They have tons of little cores integrated onto a single chips, and amazing DB throughput.

    Share
  2. How about something like Sun’s SPARC4 architecture, or even SPARC5 due out next year? They have tons of little cores integrated onto a single chip, and amazing DB throughput.

    Share
    1. Justin List Tuesday, July 3, 2012

      Your confusing old and slow cores with modern architecture. SPARC is dead.

      Share
  3. “Tilera has a the first non-x86 64-bit based server available in the market”

    So, they’re selling early-1990′s DEC AlphaServers?

    Share
    1. oh snap. You’re right. I forgot my history for a moment there.

      Share
  4. Mark Sandstrom Tuesday, July 3, 2012

    The key question shouldn’t be so much what type of cores are “best”, but how they work together. The manycore chip fabric and hw automation of the traditional system software functions are critical to scalability and efficiency going forward. Ideally, there should be application load and type adaptive manycore fabric.

    Share
    1. I agree, but first we have to get to a point where alternative architectures are deployed and tested by trusted names.

      Share
  5. One cannot ignore the Chinese in this situation, although the contenders put forward above are all viable and I like Marvell, the Chinese are producing volumes of MIPS-based processors which could fit the bill. Unlikely but still.

    Share
    1. Which vendors?

      Share
      1. The Loongson chip has been around a little while, EEtimes has many mentions of Chinese MIPS. Also take a look here for a roundup: http://www.extremetech.com/computing/127791-china-plans-national-unified-cpu-architecture

        Share

Comments have been disabled for this post