Shared bus versus switched fabric technologies

Celebrating its 20 year anniversary this year, the VMEbus is perhaps surprisingly still going strong. Approximately 38% of the embedded systems market, VME is particularly strong in mil/aero, industrial control, and instrumentation applications. VME is an excellent bus for control applications, with strength in tightly-coupled multi-processing. The Eurocard form factor and DIN connector outlay is very rugged, and the backplanes slot counts can reach as high as 21 slots in a 19- inch rack.

Right now, there is not a dire outcry in Industrial Control and Instrumentation applications for the significant bandwidth increases that fabrics offer at this time. And, while there are inherent reliability advantages in the new fabrics that can benefit nearly all applications, from a cost and compatibility standpoint, many VME applications are doing just fine — for now.

And, in the military/aerospace, many applications there are not screaming for bandwidth. But those that that need higher performance have taken advantage of shared-bus upgrades. Advances like the VME64x increased backplane performance (which is the traditional bottleneck in the system) from 80 Mbytes/s speeds to 160 Mbytes/sec, and the VME320 which is fully compatible to the ANSI/VITA 1.1- 1997 specification of VME64x has hit rates of nearly 700 Mbytes/se c with 2eSST technology (two edge source synchronous transfer).

Of course, there are communications-based and other applications in Mil/Aero that have run out of bandwidth with VME. These applications - those that are based on open technologies -- build upon existing or well-known technologies, and preserve hardware and software investments have a strong chance of doing well.

These include technologies like VME320, Raceway and StarFabric in its compatibility with cPCI (and now VME with VITA 41.5). Moreover, the VME Renaissance continues to push VME into higher performance. With Motorola's Tempe chip for the 2eSST protocol, and push in the VXS backplane (VITA 41), there is a continuing drive to preserve VME investments while boosting performance.

The VXS backplane will allow fabrics such as StarFabric, InfiniBand, Ethernet, and RapidIO to run across the P0 portion of a "new" VME64x-type backplane. Standard VME64x parallel bus can run across P1 and P2, with new switch cards runn ing the fabric across the P0 high-speed plane.

How will PCI do in this new market environment? PCI is also often used in industrial applications, and like VME will do fine in certain areas. In applications where more bandwidth is needed, advances like PCI-X bring the speeds up to 133 MHz. For telecom applications, CompactPCI, has been a mainstay for open-standards embedded applications.

With 8-slots at 33 MHz or 5-slots at 66 MHz (not including bridging), the speeds, higher pin counts, hot-swappability, and other features made cPCI quite attractive. From it's beginnings in 1997, CompactPCI had rapid growth rates until the telecom meltdown over the last couple of years. Telecom/datacom was by far the most common application for CompactPCI, and its also the application with the most demanding bandwidth criteria.

To alleviate this problem the PICMG 2.16 (Compact Packet Switching Backplane) using Ethernet and PICMG 2.17 (StarFabric) boost performance with the inherent reliability and speeds of switched fabrics. But, CompactPCI has begun to grow in other applications as well.

Growing interest
Some military communications and medical applications use cPCI, because there are specific imaging and communications needs which require higher bandwidths. And, the architecture is still doing well in certain applications in the US and has remained relatively stable in Europe and Asia. But, as fabric usage increases, standard cPCI usage to decline.

New efforts such as PCI Express and the AdvancedTCA (PICMG 3.x) architecture promise even greater performance. PCI Express is widely heralded as the de facto replacement for PCI. AdvancedTCA is geared directly for the telecom central office. With 280mm x 8U cards and wider (1.2-inch) slot spacing, ATCA allows more components on the server blades and promises Gigabit/Terabit performance.

But, how costly are the switched fabrics going to be? Looked at from an overall system level, perhaps not as bad as you wo uld think. Technologies like PICMG 2.16 and PICMG 2.17 preserve hardware and software investments and are full compatible to CompactPCI. New switch cards are required and optional node cards have been developed with the goals of keeping costs low. The backplanes have been developed and are reasonably priced.

These technologies can be used in some standard chassis that are available today, including ones with redundant cooling, N+1 power supply options, and system management. Also available for StarFabric (PICMG 2.17) are Adapter Cards and PMC modules. The Adapter Cards take a standard cPCI board, serializes the traffic and sends it across the backplane in two 2.5 Gbit/second StarFabric links.

What alternative technologies are out on the horizon to meet future demands? As processing power increases, systems will need to dissapate higher and higher wattages in smaller amounts of space. VITA 34 is an effort to develop a high-performance system in a range of applications. The new 220mm x 4U, 8U and maybe now 6U cards may be encased in metal, which allows advanced shielding and provides a housing for liquid cooling.

In the future, liquid cooling may be the most feasible way to dissipate tremendous heat build-up. StarFabric has announced a migration path to PCI Express Advanced Switching. One could start developing a system with 2.5 Gbit/second per link StarFabric today and migrate to 10 Gbit/sec Advanced Switching in 2004.