Monday, March 31, 2008

Infiniband eyes million-unit market

The pace of growth for Infiniband is on the rise, but so far the interconnect is not gaining much traction in mainstream business applications. That's the conclusion from an updated report from market watchers at International Data Corp.

IDC now expects Infiniband sales on end devices will rise from about 500,000 adapters this year to more than a million in 2011 with revenues jumping from about $125 million to about $275 million. Sales of switch ports will increase from about 600,000 this year to more than one million by 2010 with revenues on that side rising from about $250 million to about $400 million. High performance computers (HPC), database systems running on clusters and financial systems "with HPC-like characteristics" continue to be the drivers, according to IDC.

Mellanox, the sole maker of Infiniband chips, notes that the shift to 20 Gbit/s links is moving quickly with 40 Gbit/s products on the near horizon for this year. "In our last quarter, 79 percent of Mellanox's business was for [20 Gbit/s] DDR products," said Thad Omura, vice president of product marketing at Mellanox.

The chip company naturally wants to drive Infiniband into more mainstream data centers. Omura pointed to three users employing the interconnect as a cost-saving way to consolidate multiple Gbit Ethernet links on to one IB cable, a kind of unified networking approach Sun said it will drive starting later this year probably with the rollout of 40 Gbit products.

Support for Infiniband on VMware 3.5 and for NFS over RDMA in Linux could help grease the way for more such deployments. But for the foreseeable future most users will choose Ethernet if they are cost sensitive rather than a technology upgrade to Infiniband.

1 comment:

Anonymous said...

Analyst reports suffer from either focusing on CAGR or on units shipped rather than showing both and then comparing those with competing technologies, in this case 10 GbE (not GbE which seems to be the only thing people can focus on when doing any analysis these days.

It should also be obvious that moving to 20 Gbps technology is rather trivial from a technical and economic perspective. The same argument as to why people moved from 2G Fibre Channel to 4G Fibre Channel was because it either cost the same or less so reduce the other non-capital cost structures and make the move. The hard question is whether moving from 20 to 40 will follow the same path as the move from 5 to 10 Gbps signaling is not trivial and the costs may not translate.

The ecosystem being built around IB is impressive but it remains in its infancy. The IB management space is largely proprietary if a customer wants something with strong quality. Will they be willing to pay the high prices for software even if the hardware is cheap? Or will they see the modest price differential between 10 GbE and 40 IB as really a net savings in the end.