Intel lays out plans for the hyperscale datacentre

Intel lays out plans for the hyperscale datacentre

Summary: Intel has revealed what technologies it has in store to serve the hyperscale datacentre at the Intel Developer Forum in Beijing.


Intel has revealed plans to develop a reference design for a rack-scale architecture designed to serve the hyperscale datacentres.

Hyperscale tasks are lightweight computing tasks that need to be carried out in very large numbers, such as serving content in a popular web page or app.

Traditional rack servers with power hungry CPUs are overkill for these simple tasks and are a poor fit for delivering these tasks at scale because their upfront and running costs rapidly escalate.

Intel's reference rack system design will be modular at a subsystem level, allowing storage, CPU, memory and networking to be replaced independently, said Diane Bryant, Intel's senior vice president and general manager of the datacenter and connected systems group at the Intel Developer Forum in Beijing.

The Intel Atom Centerton S1200. Image: Intel

The idea behind this pick and choose approach is to allow system design to be matched to the demands of the specific hyperscale workload. In this way extraneous components that add to purchase and running costs of the system can be removed and the system can scale more effectively.

Because these systems are typically based around small form factor microservers running wimpy core CPUs with shared power, cooling and networking infrastructure they typically also take up less space in the datacentre.

The reference hyperscale design will be based around current and upcoming Intel technologies. Intel has already released the sub-10w Atom Centerton S1200 targeted at the microserver market and later this year will release the 22nm Avoton SoC with integrated Ethernet and improved performance per watt over Centerton processors. Avoton samples are with customers and the first systems are expected to be available in the second half of 2013.

Intel is working on silicon photonic interconnects and cable designs that will help with the modular design. These are intended to enable 100Gbps interconnects that have 'such low latency' that components that previously needed to be bound to the same motherboard can be spread out within a rack.

One implementation of Intel's rack scale reference architecture is already taking place in China, as Alibaba, Baidu, Tencent and China Telecom are collaborating with Intel on Project Scorpio, an initiative that will deliver a physical aggregated rack consisting of the fan and power supplies within the rack, with the goal of demonstrating TCO savings.

The Intel announcement closely follows HP's release on Monday of its new family of Moonshot microservers targeted at the same hyperscale market.

Hyperscale tasks feature heavily in the workloads of web giants like Facebook and Google, and analysts predict that microservers market will account for about one-fifth of server sales by 2015/16.

Intel also announced several new low power system-on-a-chips targeted at the datacentre at the IDF, the S12x9 product family for storage servers, the Intel Xeon processor E3v3 product family, the Intel Xeon processor E7v2 product family and Intel Atom processor for communication and networking devices codenamed "Rangeley".

The S12x9 product family shares several features with the Intel Atom S1200 processor product family, but contains technologies specifically geared for storage devices.

These include:

  • Up to 40 lanes of integrated PCIe 2.0, or physical paths between I/O and the processor to allow the capacity demands of multiple devices can be handled more efficiently.
  • Intel provides hardware RAID storage acceleration so that the computationally intensive RAID function is offloaded in hardware, thus freeing up the SoC to execute other software applications.
  • PCIe Non Transparent Bridge to provide failover support.
  • Asynchronous DRAM Self-Refresh that can protect critical DRAM data in the event of a power interruption.
  • Native Dual-Casting to allow data to be read from a source and delivered to two memory locations simultaneously, which can increase RAID-5/6 bandwidth by as much as 20 percent on a 16+2 RAID 6 system, as compared to system bandwidth without the Dual Cast feature.

Several OEMs are supporting the Intel Atom S12x9 family, including MacroSAN, Accusys, Qsan and Qnap.

The 22nm Rangeley processor aims aims to provide an energy-efficient mechanism for processing communication workloads and is targeted for entry level to mid-range routers, switches and security appliances. Rangeley is targeted to be available in second half of 2013.

Topics: Data Centers, Hardware, Intel


Nick Heath is chief reporter for TechRepublic UK. He writes about the technology that IT-decision makers need to know about, and the latest happenings in the European tech scene.

Kick off your day with ZDNet's daily email newsletter. It's the freshest tech news and opinion, served hot. Get it.


Log in or register to join the discussion
  • This never made sense to me

    You want to scale up so you use a wimpy CPU as the basis? Wouldn't the aggregate computing power of a single "not wimpy" CPU consume less than a group of wimpy ones? Wasn't this the driver for virtualization?
    • Whimpy CPUs

      This is all pretty high level stuff and doesn't take into account the real world but here is some fun with numbers.

      Xeon® E7-4870: 10 cores, 20 threads, 130W TDP, 2.4GHz clock
      Atom™ S1260: 2 cores, 4 threads, 8.5W TDP, 2GHz clock

      Multiple the Atom numbers by 15 and you get:
      30 cores, 60 threads, 127.5W TDP

      At any point in time a beefy Xeon can only do 20 things at once whereas 15 Atoms can do 60 things at once. In order to match the threads of the Xeon you are looking at 5 Atom CPUs at only 42.5W. Obviously there is motherboard power and the other components for 5 Atoms rather than one Xeon (assuming one CPU per motherboard) but consider that an Atom only has one memory channel compared to four for the Xeon, saving some power.
      • Also, it depends on the loads.

        If you're doing fewer transactions, but each transaction is more resource-intensive, the Xeon may work better. However, servers tend to do less intensive transactions in large numbers, which is where the Atom might be able to get more done, at least watt for watt.
        Third of Five
    • clock speed physics

      power consumption is proportional to square of clock speed. Thus 2 x 1GHz cores have same performance as 1 x 2GHz core but half the power consumption.

      Hence the future is in extracting more parallelism from programming. Unfortunately this is not trivial.
    • Re: You want to scale up so you use a wimpy CPU as the basis?

      Yup. Server loads are not all CPU-limited, but everything is increasingly becoming power-limited.