Industry News Desk
Intel Fields Atom for Microservers
It has forecast that microservers could get to be 10% of the server market by 2015
By: Maureen O'Gara
Dec. 17, 2012 08:00 AM
Intel is going to try going after the data center with a brand new Atom System-on-a-Chip (SoC) that can be built into relatively cheap, high-density microservers for cloud providers.
It really rather not - it really wants to sell its high-end chips - but it has no choice. It has forecast that microservers could get to be 10% of the server market by 2015 and it will have to fight for a piece of it after losing a head start earlier this year when AMD plopped down $334 million in cash and stock for SeaMicro, a microserver start-up that already had Intel designed in.
But, given the tone in its voice this week, Intel is apparently serious about the sector, which it's blown off before for defensive purposes.
Intel says the new 22nm dingus, code-named Centerton and seemingly in development since 2007, is the first low-power 64-bit dual-core SoC for these data center systems that's in production and shipping to customers.
Intel makes the production and shipping point because it's looking over its shoulder at ARM, which is promising to deliver a four-core 64-bit version of its widget for microservers by 2014. Like Intel says there's currently no enterprise-class ARM-based server chip but just wait. The ARM contingent is in major test sites.
ARM vendors have trouble buying the Centerton as a real server chip since it lacks on-chip management, I/O, networking and fabric.
Intel's part sips an un-Intel-like 6W of power - which sounds low to Intel camp followers but it's still hot and therefore expensive by ARM standards. It delivers four threads with Intel Hyper-threading.
It's also got familiar server features like Error-correcting Code (ECC) memory support for higher reliability and Intel Virtualization technology for enhanced workload management. (It's suspected that Atom always had ECC and virtualization but Intel turned the features off in earlier generations.)
Microservers, which could be sold in the droves, are supposed to be good at un-intensive compute chores like serving up web pages, content delivery, large distributed memory caching, simple Big Data search systems and MapReduce apps. Within reason, the Centerton is supposed to run the x86 server-class software data centers are used to, which ARM can't do at all.
It's unclear how many nodes Centerton can support. It pretty much depends on how the OEMs finagle the networking. Rival Calexda, which has got ARM-based microservers out for test at major accounts, say it can theoretically support 4,000 nodes and practically support 500-1,000.
See, it takes a lot of systems to process huge numbers of smaller workloads while keeping the power consumption down and such workloads can run many small but highly parallel chunks of code.
Officially designated the S1200, the Intel widget is also expected to be used in storage and networking systems and Intel says - without indicating who's doing what - that the part's got more than 20 low-power server and storage and networking systems design-wins at Dell, HP, Huawei, Inspur, Quanta, Wiwynn, CETC, Supermicro, Accusys, Microsan, Qsan and Qnap.
In fact an unnamed storage vendor reportedly swapped out an ARM design for the Intel SoC and ARMs are supposed to be pretty darn good in storage applications.
HP, which is already in bed with Calexa and its ARM-based boxes as part of its processor-agnostic Project Moonshot, means to try the Intel part in a hush-hush server dubbed Gemini.
This summer HP said the first Moonshot servers would be based on Centerton, with initial systems shipping by the end of this year. It's now more likely to be in the first quarter.
Dell's been partnering with Marvell to create so-called Copper servers using Marvell's ARM-based Armada XP chip but - since Marvell has gone dark about its development - Dell may be closer to selling Calexa boxes.
SeaMicro, the microserver pioneer that AMD had the temerity to buy - considering all of SeaMicro's gear is based on Intel parts - even Intel parts made especially for it - and will be until it switches over to ARM - has a so-called supercompute fabric that connects thousands of processor cores, memory, storage and input/output traffic and supports multiple processor instruction sets.
Calexda, which is hobbled by the fact that it's neither x86 or 64-bit, useful propaganda points for Intel though in the final analysis it may not matter, has fabric, I/O and management built into its chip.
Apparently OEMs will have to wait until later this year or early next when Intel's supposed to deliver a next-generation Avoton Atom that could make the ARM boys sweat.
It'll be built using Intel's fancy new 22nm 3D Tri-gate transistors and should have 16GB-32GB of memory and four or eight cores.
By then Intel might have a fabric too.
Karl Freund, Calexda's VP of marketing, sent around a message about the Centerton saying, "Intel didn't specify the additional chips required to deliver a real ‘server-class' solution like Calxeda's, but our analysis indicates this could add at least 10 additional watts plus the cost. That would imply the real comparison between ECX and S1200 is 3.8 vs 16 watts, so roughly 3-4 times more power for Intel's new S1200. And again comparing two cores to four, internal Calxeda benchmarks indicate that Calxeda's four cores and larger cache deliver 50% more performance compared to the two hyper-threaded Atom cores. This translates to a Calxeda advantage of 4.5 to six times better performance per watt, depending on the nature of the application."
He provided this chart to make the comparison plain:
The new Intel S1200 product family will consist of three processors with frequency ranging from 1.6GHz to 2GHz. They start at $54 in quantities of 1,000.
Despite the design-win parade Intel didn't show off any boxes so competitors figure it won't really have the chip for a while. Microsoft and Facebook are supposed to fancy the widget but it's unclear if they're using it.
Atom SoC configuration in a highly dense rack will reportedly net more revenue than a rack of way fewer, more powerful Xeon processors.
In 2014 Intel will move to a 14nm process first for low-power Xeons and then Atoms.
Reader Feedback: Page 1 of 1
Latest Cloud Developer Stories
Subscribe to the World's Most Powerful Newsletters
Subscribe to Our Rss Feeds & Get Your SYS-CON News Live!
SYS-CON Featured Whitepapers
Most Read This Week