To: Allen Benn who wrote (3833 ) 1/15/1999 8:26:00 AM From: w2j2 Respond to of 10309
Negative I2O News: Server vendors stake out turf in I/O battle eet.com :80/story/OEG19990114S0019 RESEARCH TRIANGE PARK, N.C. — The freshly christened alliance of Compaq Computer, IBM and Hewlett-Packard was scrambling this week to hammer out the first technical details of its plan to draft a specification for a new high-end serial I/O architecture for PC servers. The server companies' bid, dubbed Future I/O, has garnered an initial development partner in Adaptec Inc. (Milpitas, Calif.) as it braces to compete with an Intel-led initiative dubbed Next-Generation I/O. Both Future I/O and NGI/O seek to define a low-latency, message-passing serial architecture to handle server I/O at 1 Gbyte/second and up in PC servers that could ship early in 2001. However, the two camps differ in some key technical details as well as in their approach to royalties and intellectual property. One chief technical difference in the Future I/O spec is that, unlike NGI/O, it will accommodate PCI bus adapters. "We want to see a coexistence of PCI-X and Future I/O," said Tom Bradicich, director of architecture and technology for IBM Corp.'s PC server group, referring to an extension to the PCI bus that Compaq, IBM and HP detailed late last year. Following that decision, Future I/O advocates are considering self-timed versions of Scalable Coherent Interface or HiPPI 6400 as a link-layer interface. NGI/O, by contrast, uses a link layer based on Fibre Channel or Gigabit Ethernet-style connections. "We are looking at the trade-offs now," said Greg Still, lead architect on Future I/O for IBM. "The link layer may still be up for discussion at the developer's conference," a sort of coming-out party planned for Future I/O Feb. 11-12 in Monterey, Calif. The Future I/O link layer parallels work on the IEEE P2100 interface, which some developers describe as a channel I/O architecture layered on top of a shared-memory bus, supporting PCI-like read/write transactions. "That approach has some reliability problems," said one source familiar with P2100. "If a host issues a read/write transaction and an interconnect fails, there may be no way of recovering data." No I2O planned In its software architecture, Future I/O, like Intel's NGI/O, plans to adopt and potentially extend the Virtual Interface Architecture specification developed by Compaq, Intel and others. However, the server makers will not use the Intelligent I/O (I2O) technology developed by Intel to define a universal I/O driver model. "I2O is beyond our scope," said Still. "We don't want to dictate how to write device drivers." I2O is not a required portion of the NGI/O spec, according to one NGI/O backer who asked not to be named. "Intel's i960 group lobbied hard for using I2O, but it's extra overhead," he said. "We are going to use native NGI/O drivers rather than run I2O on top of NGI/O." Among other details, the Future I/O group is planning to deliver a full-duplex channel capable of maximum throughput of 1 Gbyte/s over 5-to-10-meter copper connections, with fiber links specified later. The group hopes to finish the spec this year, with servers using it targeted to ship in early 2001. But the Future I/O group still presents a sketchy picture. For instance, it has yet to define its basic switching fabric or to define targets for latency. "We are debating [latency targets] right now," said Still. Business arrangements, not technical differences, have been the main sticking point dividing the groups. Intel said NGI/O must remain royalty free. Bradicich of IBM said the Future I/O group is not yet sure how or whether it will charge royalties. ------ Previous