Intel preps its interconnects for next-gen supercomputers

Although it has not provided details yet, Intel Corp. has disclosed plans to develop a next-generation interconnect. The next-gen interconnect will be integrated into future Xeon and Atom server processors addressing a range of uses from supercomputers to microservers.

The news is timed to compete with Advanced Micro Devices' recent announcement about the Freedom Fabric it acquired with startup SeaMicro. AMD is expected to make it's interconnect an industry standard for its server CPUs and some emerging ARM-based server SoCs.

Sometimes called fabrics, interconnects have been key to linking tens to hundreds of thousands of processors in a broad range of servers. They are used in everything from the world's most powerful supercomputers to large cloud computing data centers and microservers, dense chassis packed with processors generally used as Web servers.

The industry has been anticipating an Intel move into interconnects after three recent acquisitions. In April, Intel bought an interconnect group from Cray for $140 million. In January, it acquired the Infiniband chip business of QLogic for $125 million, and in July 2011 it bought Fulcrum for an undisclosed sum.

Raj Harza, general manager for Intel's technical computing group, stated that they are putting a salable fabric into the processor using the assets they have along with their recent acquisitions. He also added that the company believes that this integration will happen much sooner than the end of the decade, tying it to exascale supercomputers.

The new Intel interconnect aims to cover the waterfront in applications. At the low end, it could be used in microservers to link a hundred or more processors at tens of GB/second with latencies of less than 1,000ns. At the high end, it ultimately will scale up to linking hundreds of thousands of processors at hundreds of GB/second at latencies measured in tens of nanoseconds.

The interconnects will appear in Xeon, Xeon Phi and Atom processors geared for such servers. It's not clear when Intel will relesase details of the new interconnect and CPUs using it, although some details may emerge from this week's Intel Developer Forum.

"We haven't finished the road map planning," said Hazra.

Most of the work defining the interconnect appears to be going on inside Intel with the help of a few software companies and systems integrators. "We have strong software ecosystem," he said.

The technology likely will emerge as a set of new capabilities layered on to the exsiting Quick Path Interconnect Intel uses as its proprietary processor bus. The technology may use elements of existing Ethernet, Infiniband and proprietary interconnects inclouding RapidIO, now being proposed as a competing standard for ARM server SoCs.

"We've looked at all the potential solutions and their pros and cons," Hazra said.

Intel collaborated with SeaMicro on the startup's Atom- and Xeon-based server designs until AMD bought the startup earlier this year. "We know what their fabric was and it was interesting for a microserver class products line, but I have no idea where it has evolved," he said.

- Rick Merritt
??EE Times