Intel and Microsoft absquatulate! Is all hope lost for InfiniBand?

InfiniBand has been hammered lately by Intel's and Microsoft's withdrawal from active InfiniBand development. Reporters and investors have loudly questioned InfiniBand's very survival--but, frankly, that's way too dramatic for the real story. Intel's and Microsoft's withdrawals are significant because they confirm what many developers already believed--that InfiniBand's market would not be the high-volume, mid-tier server space, but in data centers as a high-speed interconnect fabric directly hooking into CPU memory. The companies who put their money on InfiniBand as a high-end data interconnect were not too worried at the twin pullouts, though others who were developing around lower-end InfiniBand roadmaps are feeling the squeeze.

InfiniBand No Longer Network-Wide Bus Replacement

InfiniBand combines silicon chips and management software to transport large amounts of data at very high speeds. It is a channel-based, switched-fabric architecture that scales from 500MB/sec to 6GB/sec per link. It's meant to counter bottlenecks between server-to-server and server-to-storage connections by residing at both the host server and the server's storage targets. InfiniBand is particularly suited to clustering environments, since current clustering technology is often built on architectures that were never created to support clustered environments. For example, Ethernet-based clusters do not prioritize traffic across complex distributed networks, and developers must work their clustering software around these limitations. InfiniBand can enable high-speed system clusters by providing fast communications between nodes and a solid infrastructure for data movement. How fast? It offers a throughput of up to 2.5GB/sec and can support up to 64,000 addressable devices. InfiniBand originally grew out of two different designs: Future I/O, developed by Compaq, IBM, and HP; and Next Generation I/O, developed by Intel, Microsoft, and Sun.

With both Intel and Microsoft championing the first instances of InfiniB and, their withdrawal does impact the InfiniBand space. It signals the fact that InfiniBand will no longer be a network-wide bus replacement, the vision long held by the two companies. Intel was working on InfiniBand chips, concentrating on putting the soupedup chips on the motherboard and selling the faster machines as clustered servers do. However, their development roadmap lost ground to the first wave of InfiniB and introductions, which followed another path entirely: InfiniBand as a high-speed interconnect fabric in busy corporate data centers and clusters, with physical and software components distributed throughout the fabric. Intel had hoped to introduce its InfiniB and vision first in order to identify the market with its name. With market buy-in solidly behind it, Intel could afford to continue developing an InfiniBandenabled motherboard. Failing that, Intel stopped production and turned back to its 3G10 efforts, now renamed P CI Express.

Startups Smarting

From the beginning, Intel had preferred the lower-speed lx version of InfiniBand, although most other developers preferred a higher-speed 4x transport. Intel's idea was to replace the old PCI bus with InfiniBand, but that old technology stumbled back to Life with updated PCI-X and PCI Express. PCI Express is focused on signals that are contained on a single system .n a single enclosure. It focuses on signals that cross a span f inches between the processor and the nearest components. (The Hypertransport architecture is more its competitor than InfiniBand.) Intel hopes that PCI Express would do what Intel had hoped InfiniBand would do: allow it to soup up its chips to create more powerful microprocessors for Intel-based clusters.

Although Intel insists it still supports InfiniBand (it even uses the word "wholeheartedly") its move did not inspire confidence in technology investors. When Intel canceled its chip-development program, many investors felt they were left holding a $750 million bag: money already invested in InfiniBand startups like Mellanox, Banderacom and Lanel5 Software. However, the Intel move hardly spells doom for all companies, especially for those who are already funded through product completion and who had targeted their InfiniBand development efforts towards high-end data centers. Many of those investors felt doubly shaken at Microsoft's withdrawal announcement, which followed Intel's. Microsoft reported that they preferred to concentrate on development dollars on Gigabit Ethernet, which the company says addresses a higher range of server capabilities without the additional software and management expenses of InfiniBand. As with Intel, the impact is not as great as it might appear-third party InfiniBand support wil l be perfectly welcome in Windows and .NET environments.

Chris Wildermuth, director of strategic marketing at JNI, commented that Intel's withdrawal will not impact JNI's development at all, which was already centered around high-end data center interconnects. JNI's interest is "different from what the conventional interest is. When we first announced our products, we saw that InfiniBand is a data center product serving clustering." InfiniBand retains great opportunities, but perhaps not the opportunities Intel envisioned when they got into the business. Wildermuthadded, "Intel really saw this as a method to create a new connection environment directly to their process, to allow better scaling and clustering to occur. They saw this as a successor to the PCI bus."

The startups that bought into Intel's vision are suffering. Intel was primarily targeting low-and mid-ranged servers, where the sales margin was already razor thin. But this market is highly price conscious with thin profit margins, and adding fast InfiniB and chips to the microprocessors would add another $300 or so to the manufacturing cost. This cost would have to be passed on the customer to make any profit at all, and would probably cancel any performance gains the chip might offer. Some market segments would have accepted the price for the performance, but not in the volume Intel needed for its desired profit margins on the lower cost servers.

In high-end corporate data centers, the emphasis is much less on price than performance, and InfiniBand fabric products will likely fit nicely into that market. Intel never served the enterprise data center in any volume. Among the Big Iron enterprise manufacturers--Sun, IBM, HP/Compaq and Dell--Sun wouldn't use Intel microprocessors on a bet, HP/Compaq's merger delayed their and program, and IBM had already made its own deals. Wildermuth added, "If I were Intel I'd pull out too. That's no implication to InfiniB and, but where we think it's going, and where we think that analysts have adjusted their sights, they agree with our models. We don't think it has an effect on the market as we see it or as it's currently evolving. But we're not a startup whose funding was founded on its being all pervasive or all encompassing."

Michael Callahan, CTO and co-founder of Polyserve, agreed. "The reason Intel made this decision, as much as anything, is their efforts were not succeeding. They did not think what they were coming up with was as competitive as what other companies were building." According to Callahan, Intel's withdrawal of its InfiniBand hardware effort acknowledged that InfiniBand chips were not going to be coming principally from Intel in the first place. He said, "At the highest level there's been a lot of attention paid to the announcement that Intel would stop doing chip development around InfiniBand. In fact people were realizing that other companies were getting ahead of Intel quite significantly in terms of where their development efforts were leading."

Callahan, whose company develops clustering management software, believes that Intel will play an important part in corporate data centers, though not in their original vision of InfiniBand-enabled Intel microprocessors. He pointed to a wholesale movement in the corporate data center to move from UNIX systems with their fewer large and sophisticated workstations, towards Intel-based architectures running over Ethernet. If administrators use cluster management software to manage the server cluster as a single server, they can realize tremendous cost savings in the management function as well. Meanwhile, as InfiniB and enters the data center it can improve performance enormously among these clustered Intel-based servers. Callahan said, "InfiniB and plays into that because it is very useful in building a very high performance connection between multiple servers."

High-end data center managers are looking at InfiniBand with interest, not only because of its speed but because it allows clustering and storage upgrades independent of the host systems, which allows IT to separate server and storage purchases. The major software vendors in the data centers are looking at InfiniB and with interest, including IBM's DB2 division and Oracle. Still, even a high-end InfiniB and fabric must duke it out with Fibre Channel fabrics that are already ubiquitous in storage area networks, and with IP and the arrival of 10GB Ethernet. While investors in high-end InfiniBand start-ups can start breathing again, they shouldn't pop the champagne just yet.