Intel exits, but InfiniBand intact

The semiconductor giant abandons InfiniBand, deciding to apply its development budget elsewhere. But Lee Schlesinger says the decision is more about Intel's resources and less about the end of InfiniBand.

The news is grim--or is it? Semiconductor giant Intel, which had planned to develop chips to enable InfiniBand, decided to apply its development budget elsewhere. Is this the beginning of the end for InfiniBand, stillborn before it ever reaches commercial deployment?

Hardly. The announcement is more about Intel's resources than InfiniBand's promise.

InfiniBand is an interconnect fabric that can connect disk arrays, storage area networks, local area networks, servers, and server clusters. It differs from technologies built on PCI in the same way that ATM differs from Ethernet: InfiniBand uses a point-to-point, channel-based connection model instead of a shared bus. Its 10Gbps full-duplex (20Gbps aggregate) theoretical throughput is well-suited for high-end enterprise usage where a lot of storage, network, and CPU resources need to work together.

The InfiniBand market is very active, as evidenced by the 180 members of the InfiniBand Trade Association. Mellanox Technologies and IBM already have chips out in the hands of component manufacturers. Volume production is expected next year. InfiniBand will be deployed initially on high-end database hardware, processing huge amounts of stored data.

Like everyone else in the technology sector, Intel is facing tough times. It has to make decisions on what products are key to its business for the future, given that it doesn't have infinite ability to fund every avenue. Presumably, that means that whatever chips most closely support its bread-and-butter CPUs are the most strategic for the company.

While dropping its development plans, Intel says it still supports the technology, and says it will introduce a server chip set with hooks to InfiniBand. It also remains invested in InfiniBand startups like Mellanox, Banderacom, and Lane15 Software.

Instead of InfiniBand, Intel is expected to devote a lot of developmental resources to PCI Express, an upcoming I/O bus specification formerly known as 3GIO. (Don't confuse PCI Express with PCI-X, the 133MHz, 64-bit, 8.5Gbps version of the PCI bus, which competes with InfiniBand as high-performance alternatives for servers and workstations. See this white paper for a high-level overview of PCI Express.) PCI Express will co-exist with and ultimately replace the PCI bus found in every Intel-based PC since the introduction of the Pentium. It offers higher bandwidth--2.5Gbps, versus PCI's 1.06Gbps, with the promise of scaling higher by adding additional serial "lanes"--and advanced features like packet prioritization and power management.

Given the mass-market size of PCI Express and its synergy with Intel CPUs, it makes sense for Intel to apply its development resources here. Look for the first PCI Express machines to make an appearance in 2004.

I'm not worried about InfiniBand's immediate future. Throughout history, the leaders in new waves of technology have been different players from the leaders in the existing dominant technologies. None of the railroad giants became airline powerhouses. No newspaper company built a radio or TV empire. And while IBM invented the ISA bus, Intel's PCI bus took over.

It remains to be seen who, if anyone, will emerge as the giant of InfiniBand. There's room for several--adapter makers and the server vendors who employ the adapters to enhance their hardware's performance. With HP, Dell, IBM, Sun, and Microsoft among the InfiniBand Trade Association's founders and supporters, it's clear the technology has broad support.

Just as the PCI bus had to fight it out with the VESA bus, InfiniBand faces competitors like PCI-X 2.0, but InfiniBand's immediate future looks bright.

Do you think Intel's decision to drop InfiniBand will hurt the technology? TalkBack below or e-mail us with your thoughts.