Switched fabrics overtake PCI and VME

July 1, 2004
Ethernet is the early leader in the race to use switched-fabric data buses, although all the leading candidates will probably find their own niches.

By Ben Ames

Ethernet is the early leader in the race to use switched-fabric data buses, although all the leading candidates will probably find their own niches.

That was the message from Joe Pavlat, a speaker at the Military & Aerospace Electronics East conference May 18 in Baltimore. Pavlat is president of PICMG, the PCI Industrial Computer Manufacturers Group, headquartered in Wakefield, Mass.

There are five popular switched-fabric technologies today, he said: Ethernet (either 100 megabit, 1 Gigabit, or 10 Gigabit), StarFabric, PCI Express with Advanced Switching, Serial RapidIO, and InfiniBand.

Pavlat would not predict a winner. "Probably Ethernet will be used wherever it can be, while the others will be used wherever they have to be," he said. "But who knows? Don't wait for a clear winner to emerge, because that will take forever."

The story began with parallel data-bus architectures like VME and PCI. Those technologies worked well for years because their standardized interfaces allowed manufacturers to build high volumes and keep prices low.

Now, processors are growing more powerful and designers are demanding faster buses, causing the industry to shift away from parallel buses to switched serial interconnects, also called switched fabrics.

The new technology represents a fundamentally different architecture, providing a point-to-point connection to any destination as needed, controlled by an intelligent switch. Pavlat said that approach is better than parallel buses for five main reasons:

  • it is more efficient because every slot you plug into a parallel system increases capacitance — switched fabrics need less power per bit;
  • it is more reliable because the failure of any single card in a parallel system could bring the whole chain down — switched fabrics can run dual-star or mesh topologies, which route traffic around failed links;
  • it is easily maintained because hot-swapping replacement parts is simple;
  • it is more scalable; and
  • it is faster.

Faced with that challenge, designers of parallel buses have been struggling to keep up, moving from the 160-megabyte-per-second speed of VME64 to 512 megabytes per second with PCI, and now 768 megabytes per second with PCI-X.

However, switch fabrics run at gigabyte speeds, and those designs have nearly run out of slots. "So PCI-X is probably the last parallel bus we'll see, at least in the PCI world," Pavlat said.

Meanwhile, VME manufacturers have also been preparing for the change to switch fabrics, said Rich Jaenicke, director of product management for Mercury Computer Systems in Chelmsford, Mass. He spoke as a member of VITA, the VME International Trade Association in Fountain Hills, Ariz.

The primary challenge for designers of data buses is that their growth lags far behind processor performance, he said. Since 1980, processors have doubled in performance every 18 months, while personal computer bus performance has doubled every three years. And with its rigorous military requirements, the VME bus has doubled its speed every 7.5 years.

Engineers have added some switch fabrics to VME, but the latest interconnects need a new backplane to handle their multigigabit speed. That is an extra challenge for VME designers who are trying to maintain backward-compatibility between boards and backplanes.

Therefore, users are left with the question of upgrading from VME64 to VITA 41 now, and then upgrading again to VITA 46 in the future, or waiting to upgrade once when VITA 46 arrives, Jaenicke said.

Voice your opinion!

To join the conversation, and become an exclusive member of Military Aerospace, create an account today!