The winds of new technology are shifting the sands of the bus and board landscape, and the pattern is clearly serial. Parallel bus and backplane technologies like PCI, CompactPCI, and VME64 are far from extinct, but serial bus solutions are affecting even these dependable platforms.
Leading the charge is the Universal Serial Bus (USB), followed by Serial ATA for a change in storage. PCI Express will have the greatest impact, though. Boards are the centerpiece of all systems (see the figure). But getting data on and off is a requirement regardless of the product, whether through a backplane or a peripheral bus. Newer board form factors like PCI Express and AdvancedTCA are joining established ones like PC/104 and VME.
On-board buses tended to be an extension of the backplane. But two new serial standards, HyperTransport and parallel RapidIO, are enabling standards-based, high-speed, chip-to-chip communication. PCI Express is reforming the backplane, while its future cousin, Advanced Switching, joins a host of switch fabrics that are the basis for large, fault-tolerant systems.
The tables that decorate this article provide details about these technologies as well as Web links to standards and support organizations. To get the big picture, we'll take a quick tour aboard the serial bus.
BASE BOARDS A plethora of single-board computer (SBC) standards exists, led by the tiny PC/104 when it comes to expansion for compact SBCs. The PC104 Consortium will have an interesting challenge on its hands with all of the new technologies. USB is commonplace, but changing from a bus technology like PCI and ISA to PCI Express will take some innovation to retain PC/104's stacking ability.PCI-X boards continue to squeeze out performance at the high end, with servers and embedded applications likely being the limit in the near term. As with most serial technologies, the latest parallel performance often exceeds the initial standards. But a wide, synchronous bus becomes a liability as speeds increase. Ease of use will keep PC-style ISA and PCI boards around. That said, new PCI Express boards are changing the PC arena and will eventually affect the entire embedded market.
| PCI LIVES FOREVER |
| The parallel PCI (peripheral component interconnect) bus is slowly ceding to alternatives like HyperTransport and PCI Express. However, it's not falling off the radar. The protocol forms the base for bridges on all major serial bus solutions, with transparent PCI support being a key feature. This is all to maintain compatibility on the software side. It involves improving the hardware while leaving the underlying software unchanged. Applications eventually will be changed to take advantage of the new protocols, but for now we'll live with the benefits and liabilities of PCI. |
VME, CompactPCI, and AdvancedTCA deliver COTS (commercial off the shelf) support for a range of applications, such as communications and military. They're also the platforms for standards-based blade servers. All three now support multiple switch fabrics.
VITA (VME International Trade Association) has done a remarkable job of extending a venerable standard from both the fabric and parallel I/O point of view with tricks like adding pins around the existing connector standard. It even has a midlife kicker with improved speed on the parallel bus. CompactPCI hasn't changed its PCI bus, but it was out early with switch-fabric support. It's posting gains after the dotcom bust and definitely will give the alternatives like CompactTCA a run for their money. AdvancedTCA was designed from the ground up with switch fabrics in mind and lots of board space for heavy-duty systems. Design goals have been met, so it appears to be on the way to achieving the marketing goals as well. Now, the communications sector has to continue to invest in new hardware.
CHIP-TO-CHIP Intel brings a processor front-side bus out to a chip set that typically interfaces with peripheral buses like PCI and AGP. It's a great approach for single-processor systems, but it will lead to non-standard implementations for multiprocessing systems. It also is a good fit for custom approaches, though open-standard, chip-to-chip interconnects like HyperTransport and RapidIO are changing how boards are designed.Speeds for HyperTransport now reach an impressive 22.4 Gbytes/s. HyperTransort NUMA (non-uniform memory access) implementations push multiprocessing, and PCI bridging makes it an appealing approach to an on-board network.
Designers can benefit from Parallel RapidIO because it fits on board as well as on the backplane. Its sibling, Serial RapidIO, can be part of the picture or used by itself.
ON THE BACKPLANE Backplanes handle three types of data these days. The first is explicit I/O. The second is the standard parallel expansions bus. And, the third is the data that links to switch fabrics. The two dominant parallel buses are PCI and its 64-bit incarnation, PCI-X, and VME. ISA is still around on very old PCs and a surprising number of embedded systems. It continues because of the ease with which its interfaces can be created.PCI remains a significant backplane technology. It's also the main mezzanine card bus. This is changing as PCI Express becomes more common. PCI-X has created a niche, even though it marks the end of the PCI parallel bus development. The current top bus runs at 533 MHz or 4.3 Gbytes/s.
VME's 2eSST specification runs at 320 Mbytes/s. This speed will handle most applications easily, with switch fabrics taking on high-bandwidth chores.
| WHAT'S THIS SERIAL STUFF? |
| Serial bus is a term that's been applied to a host of technologies that are far afield from the old, multidrop RS-232/422 serial bus. None of the new serial technologies are really a bus. Point-to-point is the new approach, with switches at the intersections. Likewise, they all employ packet-oriented protocols. Some are strictly bit serial, like USB and Ethernet, while others push up performance by adding more connections, also called lanes. It just isn't as fun to say multilane packet connection system. |
USB leads the pack by a large margin. It simplifies SBC design since it requires only one or two USB host connections. IEEE 1394 on boards or embedded systems tends to be limited to applications that require IEEE 1394 devices like digital video camcorders. IEEE 1394b provides more robust network support and higher speeds. That said, it has yet to find its way into regular PC or embedded systems.
The one area of change is due to PCI Express. While PCI Express is the mainstay on the backplane, its ability to be easily routed outside the box suits it for a peripheral bus in many instances.
One example of moving PCI Express into the wilds outside PC is ExpressCard. ExpressCard is designed to replace PCMCIA cards. Typically, a card has one interface, but it can have both USB and PCI Express. ExpressCard hosts can handle both.
The primary limitation of the ExpressCards is that the PCI Express uses only a one-lane connection instead of the 32 maximum. Moving PCI Express outside the box usually will limit the number of lanes utilized. Nonetheless, this approach is still much more practical than PCI-X.
STORAGE One area where migration seems to be less of an issue is storage. Parallel interfaces like ATA (Advanced Technology Attachment) and SCSI (Small Computer Systems Interface) are quickly giving way to serial interfaces, SATA (Serial Advanced Technology Attachment), and SAS (Serial Attached SCSI).SATA's first incarnation doesn't break performance records, but its compact cabling and hot-swap capability turn it into the storage link of the future. SATA is making some impressive gains in many RAID applications, forcing many to rethink the use of SCSI and even SAS.
SAS should finally get moving this year. Its slightly faster performance compared to SATA is of little consequence. However, its SCSI functionality will be significant in replacing existing SCSI products. One saving grace versus SATA is that most SAS adapters also will support SATA drives. Likewise, a SATA cable can plug into a SAS connector.
One technology that keeps on plugging along is Fibre Channel, with speeds now reaching up to 4 Gbits/s. Large disk farms remain its primary use, limiting its interest to embedded developers.
SWITCHING FABRICS Next to PCI Express, switch fabrics are the hottest topic with respect to buses and backplanes. Fabrics will be found in higher-end systems in the near term and will eventually wend their way into midrange systems.StarFabric and Ethernet possess the longest track record, though 10-Gbit/s Ethernet is on the cutting edge. StarFabric represents the best alternative to Advanced Switching until the latter begins to break out of its paper standards and into real silicon. StarFabric's 2.5 Gbits/s is more than adequate for many applications. The big advantage is that StarFabric enables developers to use fabric-oriented programming techniques versus transparent bridging of PCI devices.
InfiniBand is rebounding with its 10-Gbit/s bandwidth, low latency, and ease of use. This comes on the heels of certain press outlets' words of foreboding upon learning that Intel moved from supporting InfiniBand to Advanced Switching. InfiniBand is now popping up in supercomputer designs and blade servers.
Gigabit Ethernet is appearing in many COTS fabric-based designs, but so is RapidIO. RapidIO's mix of compatible parallel and serial standards is a compelling reason to use this technology for on-board chip-to-chip communication, as well as off-board communication. Incorporating RapidIO interfaces on DSPs and FPGAs has made RapidIO the choice for computing farms.
Of all the bus and backplane technologies, switch fabrics are making the biggest difference when it comes to applications. It eliminates one of the biggest limitations on system design—how much can be placed on a single board. With fabrics, resources can be spread across any number of boards.
Old bus and board standards never die. They just get superseded. New standards are in place, and hardware is rolling off the assembly line.
See associated Table 1.
See associated Table 2.
See associated Table 3.
See associated Table 4.
See associated Table 5.
See associated Table 6.