Chinese search engine giant Baidu is using ARM-based servers from Marvell making it the first company to depend on servers using the cell-phone chip in a production environment. Baidu is using the new ARM servers in its cloud storage application named Baidu Pan.
ARM, which licenses its IP to a variety of chip makers, had stated its intentions to enter the data center market back in 2010, as worries about energy efficiency increased and the needs of webscale computing customers changed. While less powerful than their Intel counterparts, a cluster of lower-power ARM chips is more power efficient on a performance per watt basis and some workloads don’t even need the performance characteristics of a big Intel core.
The combination of these two trends has led to a plethora of vendors from big names like Marvell and AMD to startups such as Calxeda to license ARM’s cores with an eye toward making servers. Holding ARM back so far has been the delay in building out 64-bit capable cores (they are expected later this year) as well as a lack of enterprise software running on the ARM platform.
But given the economics of these so-called wimpy cores and the limits of using ARM cores in the enterprise server market today, the use of ARM-based servers in the storage arena is not surprising. Storage usage scenarios are perfect in many ways because they don’t need a lot of raw performance, nor do they require 64-bit capable cores.
Thus, Baidu using ARM for storage makes sense. It’s also an area where Calxeda expects to see its first production deployments sometime this year, according to a conversation I had with Karl Freund, the VP of marketing of Calxeda last December. As for
the Baidu deployment, it’s using the quad-core Armada CPU, Marvell’s storage controller, and a 10Gb Ethernet switch all integrated on a single system on a chip.
Marvell’s release says the chip firm customized the ARM servers specifically for Baidu’s cloud storage requirements, taking the concept of server customization common in webscale deployments to the chip level. Marvell says the platform is designed to increase the amount of storage for conventional 2U chassis up to 96 TB, and to lower the total cost of ownership by 25 percent, compared with previous x86-based server solutions. The end result should cut Baidu’s power in its data center by half according to the release.