Advanced Micro Devices has launched the first new products from its SeaMicro acquisition earlier this year, including a server that can be linked to a massive, 5-petabyte storage cluster for running large-scale big data and cloud computing applications.
The SM15000 server is available now with Intel's "SandyBridge" Xeon processors, AMD announced Monday. Two other models will be released in November: one based on Intel's newer "IvyBridge" processor, and another with AMD's "Piledriver" Opteron processor -- which will be the first SeaMicro server with an AMD chip.
The announcement shows that AMD is willing to keep developing SeaMicro systems using processors from Intel, its chief rival, if it thinks customers will want to buy them. "We're committed to multiple generations of Intel processors," said Andrew Feldman, the former SeaMicro CEO who leads AMD's new data center services division, at a press briefing to announce the new hardware.
Aside from the new processors, the SM15000 appears similar to the other high-density servers SeaMicro was already selling. The Opteron version starts at $139,000 and packs 64 eight-core processors in a 10U chassis, for a total of 512 cores and up to 4TB of main memory, Feldman said.
What's different is that SeaMicro has taken the proprietary network fabric it developed to link the server boards in its existing system, which it calls Freedom Fabric, and extended it outside of the chassis to connect up to 1,408 hard drives in a cluster.
The SM15000 allows for the right balance of compute, storage and network resources required by today's large-scale data centers, Feldman said.
The new server is designed for use by web giants such as Facebook and Google, who are crunching vast amounts of data to analyze customer behavior, and for large enterprises running certain cloud computing workloads. Prime applications include Hadoop and Cassandra, he said.
The new storage architecture offers the functionality of a storage area network but without the cost and complexity, according to Feldman. "What you get is the cost and simplicity of direct-attached storage, and you can create vastly different combinations and ratios of compute to storage," he said.
Nathan Brookwood, principal analyst at Insight64, said the announcement was also interesting because it shows that AMD is willing to keep selling complete server systems even when it means competing with its biggest customers, such as Hewlett-Packard and Dell.
But AMD also hopes to license SeaMicro's fabric technology to other vendors.
"We're in discussions right now ... with an entire ecosystem of potential partners," Feldman said. "The opportunity for us to cooperate with Dell and HP, among others, is large and that's something that's being actively worked right now."
IDC analyst Matthew Eastwood said the technology SeaMicro has developed may only be needed by 30 to 50 big companies worldwide, but that those companies may account for 20 percent of server spending because of their large size.
Web giants like Google and Facebook do a lot of software engineering in order to allow them to use relatively simple, low-cost storage systems, Eastwood said. "This would seem to give them the ability to go faster without having to do all that extra engineering on the software side."
AMD timed its announcement to come at the start of the Intel developer Forum in San Francisco. Earlier Monday, Intel announced that it plans to integrate its own fabric interconnect technology into future Intel processors, though it didn't give details.
Other vendors, including Applied Micro Circuits, are also building server components that integrate network components traditionally found on separate chips, with a view to reducing power consumption and improving bandwidth.
"It's my guess that over the next months and years you're going to hear a great deal about fabrics," Feldman said.
SeaMicro has said its network fabric can work with any processor type, and ARM processors may be an option in the future. But ARM's current, 32-bit Cortex A9 architecture is not suited to SeaMicro servers, according to Feldman. The Cortex A15, due in 2013 or 2014, may be more interesting, he said. "It's not clear yet."
