The Inquirer certainly thinks so.

However, I'm not so sure it makes sense to compare Infiniband to an as-yet-unannounced optical external QPI. QPI is currently a processor interconnect. CPUs, RAM, and devices connected by it are conceptually part of the same machine -- they run a single OS, for example. They are both "networks" or "fabrics" but they have very different design trade-offs.

Another widely-used bus in the system is closer to Infiniband than QPI -- PCI Express. Isn't it more likely that PCIe could take on IB? There are companies already who have solutions that use external PCI Express for cluster interconnect, but these have not gained significant market share. Why would QPI, a technology whose sweet spot is even further from Infiniband's than PCIe, be able to challenge Infiniband? It's hard to speculate without much information, but right now it doesn't seem likely to me.

The other prediction made in the article is that Intel's 10GbE iWARP card could squeeze IB on the low end, due to its greater compatibility and lower cost.

It's definitely never a good idea to bet against Ethernet when it comes to mass-market, commodity networking. Ethernet will win. 10GbE will win. But, there are now two competing ways to implement the low-latency RDMA Verbs interface on top of Ethernet. iWARP is essentially RDMA over TCP/IP over Ethernet. The new alternative is IBoE (Infiniband over Ethernet, aka RoCEE, aka "Rocky"). This encapsulates the IB packet protocol directly in the Ethernet frame. It loses the layer 3 routability of iWARP, but better maintains software compatibility with existing apps that use IB, and is simpler to implement in both software and hardware. iWARP has a substantial head start, but I believe that IBoE silicon will eventually be cheaper, and more likely to be implemented in commodity Ethernet hardware.

I think IBoE is going to take low-end market share from traditional IB, but I think this is a situation IB hardware vendors have no problem accepting. Commoditized IBoE NICs invite greater use of RDMA features, and when higher performance is needed, customers can upgrade to "real" IB, maintaining IB's justification for higher prices. (IB max interconnect speeds have historically been 2-4x higher than Ethernet, and I don't see that changing.)

(ObDisclosure: My current employer now sells IB hardware. I previously also worked at Intel. My opinions are my own, duh.)