New I/O bus standards like Intel's Arapahoe and AMD's Hyper Transport promise to make data shuffling between processor chips and PC peripherals faster. How do they compare? by Ong Boon Kiat

One of the most crucial yet understated metric about computing performance is the I/O bus. The peripheral connection interface (PCI) remained an immovable standard for the last ten years while processor clock speeds and memory latencies improved by leaps and bounds. Now it looks like the industry will soon receive a major bus upgrade.

Not surprisingly, the two key, next-generation bus technologies are being led by chip giants and rivals Intel and AMD. Probably because of this, the media has touted the developments a platform war that will splinter the industry into two.

Intel's offering is the 3GIO, recently dubbed Arapahoe. Based on the PCI standard, Arapahoe has the backing of key PC makers like Dell, Compaq, IBM, Microsoft and Intel. The AMD-led technology is called Hyper Transport, with key supporters Apple Computer, Transmeta, Cisco and NVIDIA.

Comparing chips
Despite media hype, however, it appears that talks of PC vendors splitting ranks over I/O buses may have been greatly exaggerated. In the last few months, both Intel and AMD have loudly declared that both technologies will be friends, not rivals.

On Intel's website, Arapahoe is called a "general-purpose I/O technology optimized towards the constraints of connecting external components and device sharing". Hyper Transport, on the other hand, was alluded to having an architecture that will "solve specific application needs." In other words, Hyper Transport can conceivably tether to Arapahoe in the same I/O chain, with each operating in its own space.

In reality, it may be hard to tell if both technologies can continue to operate with such cleanly-defined boundaries as the standards take shape. Bear in mind that overlapping features in both technologies probably means that both will have the capacity to take over some of each other's function should their proponents choose to do so. For example, both Hyper Transport and Arapahoe can serve as a chip-level interconnect.

If we look at the raw specs, Hyper Transport, with an oft-quoted data rate of 12.8 Gbps, appears to overwhelm Arapahoe's 2.5 Gbps. Raw numbers, however, rarely tell the full story. Bob Gregory, director Desktop Platform Initiatives, Intel, pointed out that the ongoing comparison of specs by the media may be misleading, since the numbers may not reflect similar operating conditions.

He said: "The published Hyper Transport for I/O applications is the one to compare to Arapahoe since it is designed for I/O applications and runs at a signaling rate of 0.8 Gbps (16-bit versions will have performance of 1.6 Gbps).

"Compare this to Arapahoe, which will run at 2.5 Gbps signaling (a 16-bit version will have performance of 5 Gbps). This is different than Coherent Hyper Transport whose numbers are often quoted but is targeted as a multi-processor bus, not I/O. An important metric of bandwidth delivered is bandwidth per pin. Arapahoe is 100 Mbps/pin and Hyper Transport for I/O devices is 29 Mbps/pin."

Other differences between the two technologies are less clear-cut. One apparent difference is the industry support that both have garnered so far, which will affect the standards' eventual adoption. Not surprisingly, Gregory touted Arapahoe's openness and "evolution(ary)" approach (it is essentially PCI 3.0) as its key advantage.

But just as Intel has its usual list of impressive backers, Hyper Transport has garnered good support too-from Apple Computer, Transmeta, Cisco and NVIDIA. And according to many industry watchers, one of the reasons that many vendors chose to support AMD is precisely because they do not want to become too reliant on Intel.

Intel, of course, has another I/O standard in waiting, called Infiniband. This technology is often touted as a PCI-replacement, but for many, Infiniband's ultimate purpose is to connect servers as part of a processing cluster, which is more of a data centre application. As such, it doesn't compete with PCI or Arapahoe, which are PCI-based bus standards.

Hyper-active
It is at the other I/O scale at the silicon level that AMD envisions using Hyper Transport. This is where processors needs to be connected together both for signaling and data transfer. For example, Cisco could use Hyper Transport inside its switch to improve reliability and connection speeds between silicon devices on system boards. According to the head of the Hyper Transport consortium, Gabrielle Satori, the Hyper Transport technology is meant for implementation up to the I/O link.

But support or not, Hyper Transport is likely to face competition from Arapahoe in the chip-to-chip interconnect stakes simply because both technologies can operate there. In addition, Hyper Transport will see competition from another technology: RapidIO.

RapidIO is targeted to the embedded market, and receives firm support from Motorola and IBM. Both companies plan to use the standard in the front-side bus of new PowerPC processors that will be shipped in new RS/6000s and Apple Macintoshes. Like Hyper Transport, RapidIO is scheduled to splash into market this year. One early adopter is US-startup Red Switch, which said that it will release an eight-port RapidIO switch by the middle of this year.

Actually, both RapidIO and Hyper Transport proponents eye a far bigger market beyond the PC space. That is to provide embedded interconnect silicons between chips in communications and mobile devices. But both may have to duke it out with each other in the coming years.

As for PCI, MIS can expect to see PCI 2.3 continue to be implemented in home and corporate desktops in the next two years, before Arapahoe wades into the fray in 2004. For workstations and servers, PCI-X is expected to be the I/O staple until 2003 when Arapahoe is expected to get early adoption.

The next bus
The need for faster buses is obvious to chip makers like Intel, which saw the futility of cranking up clock speeds in their microprocessors only to have laggard PCI buses slow down its data links to critical computing components like memory, disk and video. But replacement of legacy bus technologies and finding a symbiotic existence with new rivals is as much a political issue as it is a technical issue.

For now, the pieces appear to fall into neat places. To take a simple analogy, Hyper Transport can be seen as a next-generation rail service that operates within municipals, while Arapahoe is a train service that plies between different states. Infiniband? Think of it as ultra-fast, exclusive freight trains that have the potential to link between states.

Well, at least that's the plot tacitly agreed on by all working parties so far. Time will tell if the script will change.