Buses Dominate Board Architecture
PCI is dead. Long live PCI, PCIX, CompactPCI, and so on. Of course, the emerging plethora of high-speed serial buses is looking to make mincemeat out of standardized parallel bus architectures like PCI. Fortunately for many, the emergence of high-speed serial buses will take some time. Established parallel bus-based boards will be found in the bulk of shipping products.
This year will be a watershed for real implementations of packet-based bus technology standards like InfiniBand, RapidIO, StarFabric, and HyperTransport. Even Ethernet on the backplane and Serial Advanced Technology Attachment (ATA) for hard-disk drives will make an impact this year. The big question isn't when or why, but rather, which one fits the designer's criteria?
Board-based technologies like HyperTransport will make their way into new systems. Support for existing interface technologies such as PCI, PCIX, and CompactPCI allow the incorporation of HyperTransport without major modification to support hardware. Initial implementations will take advantage of the scalability features of most serial interfaces. HyperTransport starts with 2-bit channels. Now 4- and 8-bit support is available, with 16-bit support to show up this year.
HyperTransport is going head-to-head with RapidIO. The new HyperTransport specification will match RapidIO's multimaster operation, but RapidIO keeps things simple. It also uses standard low-voltage differential-signaling (LVDS) technology, allowing standard libraries to readily incorporate RapidIO designs. This is especially true in embedded SoC applications where multiple IP is tied together in one chip. In addition, RapidIO supports nonblocking switches that provide the necessary high performance for environments like telecommunications.
StarFabric looks to make significant inroads into new PCI backplanes. While native mode StarFabric endpoints are the long-term goal, PCI bridging can be done easily, making it a great way to extend the number and length of PCI and CompactPCI systems. The technology should meet scalability requirements. Its elegant design makes StarFabric suitable at the other end of the spectrum with just a few endpoints.
InfiniBand will be coming up to the bus, but the long-term goal of native InfiniBand host interfaces will have to wait for lower-cost silicon. Still, InfiniBand is just where it wants to be with its initial use in higher-end, high-performance systems that require high reliability. InfiniBand adapters for PCI and other existing buses will make the bulk of the shipments this year, with some linkages to new interfaces like StarFabric and HyperTransport seeing the light of day.
What of 3GIO—now PCI-SIG? It's still in a state of flux, and real products aren't expected to appear this year. But it's a technology worth watching simply because of Intel's interest in it. It will have to deal with a number of established players, including HyperTransport, StarFabric, and RapidIO, though the overlap isn't complete. It will be interesting to see whether peaceful coexistence or down-and-dirty fighting will reign in nine months.
One thing that designers need to get used to is the fabric architectures that employ serial interfaces. Wide multidrop buses like PCI have no counterpart to the fabrics. Hubs and switches used in fabrics will be more familiar to network administrators. Fabrics, though, are the wave of the future. Luckily, designing systems using hubs and switches is significantly easier, especially when it comes to expansion and reliability. More systems will feature high availability in addition to scalable interconnect performance.
Serial ATA may turn out to have the least resistance as it emerges this year. Transparent to the BIOS and operating systems because of its ATA interface, Serial ATA will succeed if disk-drive vendors deliver. The simplified wiring will be a boon to all disk-drive users and a definite plus to embedded developers where peripheral wiring is always a curse. Neither the universal serial bus (USB) nor IEEE 1394 made it as internal devices, so Serial ATA has a ripe market. The Small Computer System Interface (SCSI) still has its niche, while parallel ATA is definitely running out of steam.
The cry "Ethernet everywhere!" seems true even when it comes to system backplanes. While Ethernet is going to have a rough time displacing the competition from RapidIO, StarFabric, and 3GIO, it may still wind up being a major player given the established Ethernet infrastructure and support. This probably won't be a breakthrough year for Ethernet on the backplane, but there will be more noise about it as the months roll on.
Serial ATA could have the greatest impact on storage since the introduction of the integrated development environment (IDE)—that is, if hard-disk drive vendors can crank out devices as quickly as the chip vendors deliver Serial ATA controllers. More importantly, motherboards with built-in Serial ATA support will be key to locking in this new standard. Development and diagnostic tools are finally rolling out, so this could be a good year for new hard-disk drive technology. Embedded systems may wind up being the primary target for initial Serial ATA hardware and disk drives because designers are looking for the best, least expensive, and most robust solution available. They have control over the whole design and don't have to contend with users looking for eventual upgrades of their PCs. If there's Serial ATA inside the box, users don't even have to know it's in there.
USB 2.0 looks to be tops on the desktop but there's still a lot of heat from IEEE 1394, also known as FireWire. USB 2.0's speed improvements over its predecessor provide sufficient bandwidth for almost any PC peripheral, so it will simply be a matter of convincing motherboard vendors to include USB 2.0 support on new products. This definitely isn't a hard sell. IEEE 1394 will continue to push the multimedia arena, but it remains to be seen whether or not the more ruggedized version will become dominant in auto entertainment systems. Its niche in digital camcorders may be secure unless USB 2.0 makes a dent this year.
It's an open switched-fabric standard for system interconnects that's royalty-free and can win many converts. There are some major players in the StarFabric Working Group, so it will be interesting to see the current crop of chips and software turn into real products. The initial set of products provide gateways from standard buses like PCI to StarFabric. This year, we may see more native StarFabric endpoints.
Serial buses are all the rage, yet parallel buses like CompactPCI, PC-104, and VME will remain major players this year. There's just too much support hardware available for this to change, at least this year. Besides, most developers are looking for stability, not the bleeding edge. Serial buses will continue to make inroads, but parallel buses will stay on top for now.
AMD's HyperTransport technology for processor/device-level interconnects has moved under the auspices of the HyperTransport Technology Consortium. This move gives it broader appeal and should make 2002 a watershed year for HyperTransport. Its scalability makes it ideal for a range of motherboard- and system-level architectures. Its low pin count will help in new processor development where HyperTransport support is built into processor chips. Such an architecture could greatly expand options for embedded processors by providing standardized peripheral connections.
Last year, InfiniBand chips and management software started to appear. This year, the promise of InfiniBand will finally come to fruition. Interoperability will be key, although many vendors will try to provide end-to-end solutions. InfiniBand probably won't make it to the motherboard level in the next 12 months, but falling fabric costs should push it in that direction. InfiniBand is poised to make a major impact on storage if vendors continue to deliver the goods.
Programmable logic creates customized solutions.
Will FPGAs and reconfigurable chips be the bane or bonus for board vendors? Support software will be key. Until now, boards with reconfigurable hardware required significant development expertise to configure. This year could prove interesting as the standardized libraries grow and configuration tools approach those used for board layout.
The view of customizable systems continues to change as more tools are targeted at moving boards built from standard components into a chip. A smaller platform results with the same functionality of the original. It's great for migration and a suitable path for prototyping because board designers have familiarity with building boards.
HyperTransport and StarFabric are going to give RapidIO a run for its money, but RapidIO is going to push its advantages to stay in the running. Its parallel packet-based operation is simple to implement. RapidIO's small endpoint footprint can even be implemented in an FPGA. Support for a serial version minimizes pin count where the highest performance isn't necessary. It's an "in the box" packet switch technology that's worth watching this year.
Intel's 3GIO, now under the auspices of PCI-SIG, won't ship this year, but keep an eye on its advances. Real products will begin showing up in a year or so. Its low latency, quality-of-service (QoS) attributes and high-availability support will be critical to its success. Of course, it's a high-speed serial bus. 3GIO is supposed to coexist with emerging standards like HyperTransport, InfiniBand, and StarFabric. It will be interesting to see how this is done and what 3GIO winds up displacing in the long run. PCI isn't the only thing that it can eliminate.