I profiled silicon-on-sapphire platform startup Peregrine Semiconductor last month. Now, I turn my attention to Mellanox Technologies. This company is uniquely positioned as one of the few surviving Infiniband silicon vendors.
Founded in March 1999, Mellanox is betting $89 million in venture capital on developing a suite of Infiniband products, from silicon to host adapters to switching systems. The company is now positioned as a fabless semiconductor vendor; its system offerings-adapters and switches-are not marketed directly but sold instead to OEMs as either reference designs or for resale with the OEM's branding.
To be sure, Infiniband has had a rocky couple of years. One problem: the technology was, improperly, marketed heavily for storage applications as a replacement for Fibre Channel. Though Infiniband could technically replace that longstanding storage interconnect, Fibre Channel has a large installed base, proven performance and a healthy technology road map. Another problem: When the going got tough, the tough got going-right out of the business. Both Intel and IBM Microelectronics made high-profile exits from the Infiniband silicon market.
As Infiniband emerges from the chasm, the technology is taking its rightful place in server clustering, replacing proprietary interconnects like QsNet from Quadrics; Myrinet from Myricom; SCI and Sun Fire Link from Sun; and SP Switch2 from IBM. Infiniband is not necessarily superior but its value proposition in high-performance computing and enterprise data centers lies in being part of a larger trend toward open-architecture systems. The fastest proprietary supercomputer, the NEC Earth Simulator, cost over $5 million per teraflops; by contrast, an open-architecture cluster based on Apple G5 processor blades and an Infiniband interconnect cost Virginia Tech a bargain $500,000 per teraflops.
When Intel and IBM quit the Infiniband silicon market, I'm sure Mellanox's employees and investors were both jumping for joy and quaking in their boots. Today, Infiniband is capturing a comfortable niche in the high-performance computing segment. That alone will likely not bring Mellanox to profitability.
Nonetheless, the enterprise data center, a la Oracle 10g and IBM DB2, is waiting in the wings. If Infiniband catches on there, it will be a fun 2004 for both Infiniband and Mellanox.
Jeremey Donovan (jeremey.dono-van@gartner.com) is chief analyst at Gartner Dataquest.
http://www.eet.com