InfiniBand throttles down

Electronic News, July 22, 2002 by Gale Morrison, Tom Murphy

The severe economic pressure on the computing and communications industries is putting the squeeze on the InfiniBand data I/O scheme. Talks with executives across the connectivity spectrum last week show that if InfiniBand reaches critical mass, it will happen later than expected, and the total available market will not be as large as everybody once thought.

InfiniBand uses a combination of silicon and software to move around big chunks of data. It requires that silicon and software be in place at the host, such as an array of disks containing an Oracle database, and every target, such as a backup tape system. It was engineered to replace Fibre Channel and avoid the slow and inefficient way that the omnipresent TCP/IP software protocol moves large amounts of data at great CPU processing power expense.

Since InfiniBand was introduced, however, there has been progress in rival technologies such as iSCSI, which is specifically intended for storage and works off the deployed SCSI platform and standard IP. There also has been progress in PCI Express, which provides a fat pipe to move data around inside servers, as well as in the deployment of Gigabit Ethernet and 10 Gigabit Ethernet connectivity, including currently available GigE silicon for the rock-bottom cost of 10/100 Ethernet silicon. All of these engineering developments have eaten away at InfiniBand's opportunity, especially as they reach volume pricing and are adopted more quickly.

That's not good news for the venture capitalists who have thrown nearly $750 million behind InfiniBand-centric companies such as Mellanox, Banderacom and most recently Topspin. Nor is it good news for large corporations such as IBM and Sun Microsystems, among others, which have poured perhaps $250 million into InfiniBand system engineering.

Fortunately, though, this reality is not blindsiding anyone. Executives have greatly tempered their expectations for InfiniBand's proliferation and are concentrating on using it in the heart of data center systems where it can really help the cause.

"InfiniBand a year ago was a bit more grandiose and aggressive," said Greg Dahl, marketing VP at Austin-based Banderacom, a pioneer InfiniBand silicon vendor. Dahl believes the drop-off in design activity - which everyone has felt - was most severe in Q42001, and Banderacom and others have since acclimated themselves.

"The bulk of the (downturn) really already hit InfiniBand," Dahl said. "I don't think that the deployment rate has changed much in the last six months.... We're going steady and working with our beta partners."

Kevin Deierling, VP of product marketing at Mellanox, said the InfiniBand message is resonating with IT managers because it allows for on-the-fly clustering and storage upgrades, decoupling the storage element and processing element purchases. Both the purveyors and users of systems and software for IBM's DB/2 and Oracle are said to be pushing for InfiniBand deployment.

However, many executives say the current economic environment, with so few dollars available for development or purchases, makes InfiniBand very hard to justify. And it flies in the face of computing history that shows closed and proprietary systems losing out to more ubiquitous and cheaper-to-implement standards.

One analyst who spoke on condition of anonymity echoed what many others have said privately to Electronic News in recent months.

"InfiniBand appears to be struggling because there's little value-add compared with IP as a protocol and in other applications such as a backplane protocol. It similarly adds no value," the analyst said. Others are diplomatic but still admit they have left InfiniBand as a "you-have-to-pick-your-battles" kind of decision.

"We spent a lot of time looking at InfiniBand," said Mark Aaldering, director of marketing for IP cores at Xilinx Inc. "We chose to invest in Rapid I/O and HyperTransport and PCI Express." Programmable logic vendors have found themselves smack in the center of the connectivity battle because PLDs can be used to prototype new systems with the new standards and bridge one connectivity scheme to another.

Meanwhile, Intel is putting much more emphasis on the scheme that gives it the most muscle and requires plenty of processing power, relegating InfiniBand almost solely to the data center.