NEW YORK IBM Corp.'s Microelectronics Division (East Fishkill, N.Y.) has revealed specifications for its first three chips based on the Infiniband architecture. IBM said it would also add Infiniband functionality into its ASIC core library.
The chips, which were announced this week at the Infiniband Developers Conference in Newport Beach, Calif., will be introduced in 2001, IBM said.
The Infiniband architecture is designed to speed communication between systems and enable the clustering of servers and peripherals. IBM is one of several silicon vendors developing products that support the communication architecture.
Infiniband is a switch-fabric architecture that uses four-wire links and multiple levels of redundancy to off-load processor cycles devoted to communications. An Infiniband link operates at 2.5 Gbits/second using either packet or connection-based methods.
The three Infiniband chips that IBM has described a host channel adapter (HCA); target channel adapter (TCA); and switch chip will work together to enable server clustering. The chips will be manufactured in 0.18-micron CMOS at IBM's fabrication facility in Burlington, Vt., according to Jim Bowers, server architecture marketing manager at IBM Microelectronics.
Bowers said IBM drew on its mainframe experience with switch-fabric architecture to develop the Infiniband chips, and incorporated features such as on-chip serial links and Infiniband protocol handlers.
Responding to customer interest, IBM will make its Infiniband technology part of its ASIC core library for custom chip development, Bowers said.
IBM's three standard chips will support 1x or 4x Infiniband, or up to 10 Gbits/second of serial-bit bandwidth, the company said.
The host channel adapter (HCA) will provide a PCI/PCI-X interface to enable servers, workstations and PCs to support Infiniband output. The chip allows a direct application-to-application connection over an Infinib and network, with zero copy and without context switching, IBM said.
The TCA chip enables Infiniband input for PCI/PCI-X-based peripherals such as a RAID controller, Ethernet adapter or router. The chip includes the Infiniband transport layer, management layer, network layer, link layer, physical interface, and the logic required to translate between PCI semantics and Infiniband message passing.
The switch chip, which sits between an HCA and TCA on a system motherboard, card or switch cluster, allows for packet information transfers from a host to I/O. The 8-port 4x switch supports up to five virtual lanes per port and routes packets from one port to another. The physical layer includes IBM's patented data encoding and transmission scheme for enterprise system connection.
IBM targeted Infiniband 4x rather than 1x because "we feel it is the right price/performance given the state of server bandwidth today," Bowers said. "4x is the bandwidth we feel makes Infiniband valuable in the marketplace."
IBM said its Infiniband chips will eventually come in 1x, 4x, and 12x flavors. The company said its future Infiniband chips will also be implemented in its own systems, though Bowers declined to say when Infiniband hardware will be available from IBM.
The Infiniband Trade Association now counts more than 130 companies as members. In addition to IBM, its seven member steering committee includes Intel, Hewlett-Packard, Compaq Computer, Dell Computer, Microsoft, and Sun Microsystems.