Standard I/O: Next-Generation Server Bus Standards Emerge

When the next generation of advanced PC servers goes to market next year, the existing PCI bus specification won't be adequate to provide the necessary throughput to run large-scale applications. Current Intel Xeon chips require about 800MBps of throughput, but the next generation of 32-bit Intel chips (code-named Foster) require about 3.2GBps of throughput. To compete with UNIX, the Intel architecture requires a new host bus standard.

For more than a year, OEMs and their customers have wrestled with the idea of supporting at least two competing standards for the replacement of the PCI host bus adapter. The Next Generation I/O (NGIO) and Future I/O (FIO) specifications are similar in overall design but different enough that the two groups proposing these standards had difficulty resolving the differences to create a unified standard.

In early August, the NGIO Forum (http://www.ngioforum.org), which consists of Intel, Dell, Sun Microsystems, Hitachi, NEC, and Siemens, released the 1.0 Specification for the proposed replacement of the PCI bus architecture. At its first developers' conference in September, the NGIO Forum presented the specification, which it based primarily on 2 years of Intel's work.

NGIO specified a channel-oriented, switched-fabric topology for data transfers between servers, storage systems, and other network systems. The NGIO specification would let vendors build Host Channel Adapters (HCAs), Target Channel Adapters (TCAs), switches, physical links, fabric services, and software to interoperate between different OSs.

With a transfer rate of 2.5GBps per wire (i.e., 2.5GBps multiplied by the number of wires), the NGIO specification is about 500 percent faster than PCI-66. This architecture would let bandwidth increase proportionally with the number of connected devices and provide communications between peripherals without the intervention of the host CPU and bus.

A new addition to the NGIO specification was Sun's Fat Pipes technology. Beyond sixteen 32-channel wires, NGIO became throughput limited. To provide gains from 10 to 100 times the throughput of bundled wires, Fat Pipes extends the NGIO architecture and software and provides faster interconnects between peripheral devices. Fat Pipes is in development and will be fully specified by first quarter 2000.

The other group with a host bus standard is the FIO Alliance (http://www.futureio.com), which includes IBM, Compaq, and HP. FIO is similar to NGIO, but the components provide high throughput at a high cost. According to some estimates from the NGIO Forum, a server's FIO implementation would be eight times as expensive as NGIO. However, FIO is compatible with PCI-X, and NGIO isn't. (PCI-X is a backward-compatible high-performance extension of the PCI local bus that addresses the increased I/O requirements of high-bandwidth applications. The PCI Special Interest Group—SIG— http://www.pcisig.com—is reviewing PCI-X.)

The three FIO players believed that NGIO wasn't high-powered enough to run the mission-critical applications of the future. Members of FIO also thought that the low cost of NGIO favored commodity PC manufacturers such as Dell, which don't fund the future research necessary for these types of technological advancements.

In fact, Intel and the NGIO Forum set their goals to ensure that "the bus architecture will be robust, broadly adopted, and licensed on a mutual royalty-free basis." FIO planned to create a standard that provided tight control, was proprietary, and provided a royalty stream to the group. Kimball Brown, vice president and chief analyst for Dataquest's Servers Quarterly Statistics program, said, "The Big Three want to prevent their installed base from being trashed with commodity hardware that anyone can sell. Instead of trying to lock things out, they are trying to provide a more robust standard, which makes sense to me."

Money motives aside, the FIO group wanted a standard that was open to innovation and engineering, and Intel wanted an industry standard that the company could count on to prevent compatibility problems with its next generation of chips. Brown said, "Innovation means differentiation, and what Intel wants to have is an identical product. This [uniformness] is so important in terms of getting chips, such as Foster and McKinley, up and running. Intel is more proprietary, even though you can get it from about 80 people instead of three."

Yet, sanity has prevailed, and the two groups have agreed to merge their standards into a unified standard called Standard I/O. Apparently, the groups' negotiated agreement has resolved many of the technical concerns. The new bus standard will incorporate most of the commodity features that NGIO wanted for low-end servers and will add Sun's Fat Pipes architecture for high-end servers. The emerging details suggest that we'll see a 2.5GBps bidirectional wire transfer rate with 1, 4, or 12 wires in the bus. The details also imply that the FIO Alliance's proposal to use PCI-X as a stepping-stone to the new bus architectures is dead. So, PCI-X is left to fend for itself.

The issue of innovation and differentiation was at the heart of the original disagreements that led to competing standards. With many details remaining undecided, the specification looks as though it will provide enough flexibility for innovation from vendors such as IBM, Compaq, and HP so that the companies can differentiate their products from commodity server vendors such as Dell. Still, the NGIO Forum should get its desired commodity bus standard. How all the fine points will get resolved is anyone's guess because the two standards are different enough that the Standard I/O group will have a devil of a time satisfying both sides.

To have had two standards for a PCI bus replacement wouldn't have benefited the consumer. Competing standards would have meant that fewer boards would have been available for servers, and that vendors such as Adaptec would have had to spend resources creating more variants of their products for each standard rather than innovating new hardware. So, Standard I/O is good news. IBM and Intel will chair the committee and work out the details through the spring of 2000. The result of this crossbreeding is that the future server bus won't be a horse (FIO) and it won't be a donkey (NGIO); the next generation server bus will be a mule (Standard I/O). And the mule will carry the load.