As Santayana noted, "Those who cannot remember the past are condemned to repeat it." And embedded history is about to repeat itself again. The InfiniBand powers-that-be have decided that the emerging switch-fabric technology is for mainstream servers, not lowly embedded deployment. "Embedded need not apply" is today's attitude.
Deja vu. Remember when PCI first emerged? Developed as a peripheral bus for PCs, it was ruled off-limits to embedded participants. The PCI SIG concentrated on PC slot-card, not embedded implementations. The resultant slot card fit PC desktops and servers but lacked many high-reliability features required for embedded deployment.
But embedded folks didn't give up. Ziatech and PCIMG (the PCI Manufacturers' Group) created their own PCI-based buses, like PCI/ISA and CompactPCI. These are PCI buses with different electromechanical implementations built on inexpensive PCI silicon. The PCI/ISA bus integrates PCI and ISA on a single slot card. CompactPCI mates PCI with the high-reliability Eurocard board form factor and DINN connectors.
Integrating PCI with different bus/board implementations was very successful. PCI now dominates as a universal desktop, server, and embedded bus standard, deployed at four design levels—system bus (CompactPCI, PCI/ISA), peripheral bus (PCI), mezzanine bus (PMC, PC/104 Plus, PC-MIP), and systems interconnect bus (PC-to-PCI serialized connections).
Here's the real kicker. PCI is being challenged in PCs by USB, and in PC servers by InfiniBand and Intel's new peripheral bus (or AMD's HyperTransport), at the upper and lower levels, respectively. PCI will survive, partially as a low-end add-on bus for PCs and portables, but mainly as an embedded system bus. So embedded PCI, initially spurned by the SIG, might well be PCI's savior.
History is repeating itself with InfiniBand, developed by Intel and server vendors. Like the PCI SIG before it, the InfiniBand Trade Association (IBTA) is focusing on the PC, i.e., PC servers, and avoiding embedded needs. Embedded vendors and PCIMG have been given the cold shoulder by the IBTA.
Thus, in its first incarnation, InfiniBand won't be deployable for many embedded applications. Its boards and connectors are more server-oriented than suitable for high-reliability embedded deployment. Plus, InfiniBand targets data center servers and networking, implementing embedded QoS, switching, and routing in the switch node.
InfiniBand is on the fast track for success. Sili-con—HCA, TCA, and node-switch chips—is emerging from multiple sources that will be fielding it by year's end. Many are sampling now. Also, the software layer is falling into place.
InfiniBand delivers a high-bandwidth switch fabric with multiple connections for both server clusters and system networks. InfiniBand is expandable, it supports hot-swapping, and it has high RAS. Built on a crossbar switch-node and channelized I/O, it supports higher-level switching and routing in the fabric. InfiniBand additionally provides a smart I/O channel to offload CPU I/O processing, a necessity for GHz and above RISCs
A winner today, InfiniBand will surely include embedded deployment tomorrow. History will repeat itself. Just as with PCI, the embedded folks will build their own InfiniBand . This will be a good thing for embedded server nodes and you never know when your current technology winner will fall to a faster, better alternative. Embedded deployment has proven to be a good technology life-insurance policy.