IBM is working on several advances to try to make the bricks surpass current monolithic designs. First off, because the systems are connected on all six sides to their neighbors and must easily slide in and out, they can't be connected using ordinary plugs and sockets. IBM is using a technique Garner developed called "capacitive couplers," small pads that can send signals from one brick to another with which it's in contact. A key part of the brick concept is data protection. As with current RAID (redundant array of inexpensive disks) technology, information such as a database file is stored on several disks so it won't be lost if a single disk fails. The brick idea takes that concept one step farther, with the bricks automatically shuffling data from one brick to another to compensate for problems such as failed drives, bricks or networking elements. And of course the bricks suffer an extreme version of a long-standing issue with high-end computers: overheating. In current prototypes, the bricks are impaled on vertical pipes that hold a water cooling system. Interior bricks, with no air circulation at all, would otherwise quickly suffer debilitating temperatures. "I believe the industry is moving toward liquid cooling," Garner said. Water cooling has its problems, though, including expense, maintenance difficulties, and a required connection to a specialized external system to cool the circulating water back down. IBM itself has switched its top-end mainframe servers from water cooling to air cooling. More challenges arrive when building a three-dimensional lattice of networked nodes--each storing its own data along with routing information to find other data. IBM has been researching issues such as how many units can fail before the spread of data through the system is impaired. The prototype brick system connects each brick to the other using Ethernet networks with a data transmission speed of one gigabit per second, but that version of the technology won't work for actual products because transmission delays, or latencies, are too long, Garner said. "Latencies are one of the issues with gigabit Ethernet," Garner said. Eventually, IBM expects to use either InfiniBand or 10-gigabit-per-second Ethernet.





