Currently, everything from modems to network cards plugs into computers using the widespread PCI (Peripheral Components Interconnect) standard. But increasing the speed of PCI will become prohibitively expensive, and engineers have been searching for an alternative that will let computers keep pace with ever-faster CPUs
Rival chipmakers Intel and Advanced Micro Devices each had designs that could do the trick, with AMD's HyperTransport further along than Intel's. But now, after the industry group in charge of PCI met Friday, the question of succession has been largely decided in favor of Intel's technology, code-named Arapahoe and sometimes called 3GIO. The political maneuvering was first reported by CNET News.com on Wednesday.
A final vote by PCI-SIG is scheduled for Friday, after lawyers from member companies have had a week to review Arapahoe, but the group is expected to adopt the technology, PCI-SIG President Roger Tipley said in an interview.
"The intent of Arapahoe is to be the one unifying input-output technology of the future," Tipley said. "There wasn't anything negative said about adopting the technology. The extra week of due diligence may not have been warranted, but it helps all the board members feel better."
Arapahoe could be used to speed the transfer of data to not only network cards but graphics systems as well, and to the subsystems that must funnel the huge amounts of data associated with digital images or video, said Bala Cadambi, third-generation interconnect program manager at Intel's Desktop Platform Group.
A computer overhaul this deep, though, involves more than just a routine adjustment to support a faster new CPU--something Intel has done every few months for years.
"Ultimately, the goal is to make sure we have a specification ready where products are ready by the second half of 2003," said Michelle Leyden Li, platform initiatives manager in Intel's Desktop Platform Group. Intel plans to describe Arapahoe details at its Intel Developer Forum beginning Aug. 28.
The latest version of PCI, called PCI-X, satisfies current data-transfer needs, said Instat/MDR analyst Cary Snyder, but that will change. "In the next couple years...PCI-X no longer has the bandwidth. That's when Arapahoe would be ready to come into its own," he said.
Built for speed
Though Intel and PCI-SIG executives declined to give details on Arapahoe, Tipley said the standard will offer speed, cost and longevity benefits over the current PCI while using the same control commands. That means today's PCI software won't have to be rewritten to support Arapahoe hardware.
Whereas today's PCI transfers data across 32 or 64 parallel wires, Arapahoe will use fewer, higher-speed lines, and data won't have to be synchronized across the collection.
Current PCI-X has a total bandwidth of 1.1 gigabytes per second, with each wire carrying about 138 megabits per second. When Arapahoe is ready, that may increase by a factor of four to eight, but Arapahoe will be even faster than that, Tipley said.
AMD positions its HyperTransport technology as a way for one chip to communicate with another--for example, a CPU to communicate with a memory controller or a PCI subsystem. Companies such as API NetWorks are interested in using HyperTransport as a successor to PCI, but AMD apparently is not.
"The opportunity has been available" for AMD to propose HyperTransport as a PCI successor, but they didn't approach PCI-SIG, "despite my entreaties," Tipley said.
AMD, which could not immediately be reached for comment, is a member of the PCI-SIG board of directors, as are Intel, Compaq, IBM, Microsoft, Hewlett-Packard, Texas Instruments, Phoenix Technologies and Broadcom's ServerWorks division.
But HyperTransport, now governed by the more neutral HyperTransport Technology Consortium, will live on, if not as a PCI successor. It will be used as the "front-side bus" data pathway that connects AMD's future Hammer line of CPUs to the rest of the computer. Several networking chipmakers such as MIPS and PMC-Sierra also plan to use it.
"I do believe there is a lot of momentum for HyperTransport in the communications market," Insight 64 analyst Nathan Brookwood said.
Because HyperTransport and Arapahoe/3GIO are technologically similar, political maneuvering becomes more important, Brookwood said. "There's little doubt in my mind that 3GIO is at least as good a solution as HyperTransport," he said. But, "the fact that it's positioned better in terms of the PCI-SIG would help 3GIO become a de facto standard."
Learning from the past
Wrangling a herd of competing computing giants into the same direction is tough, though, as recent experience has shown. Even extending PCI to PCI-X--a technology originally backed by IBM, HP and Compaq, with Intel only adding support later--was tough going.
Intel was given short shrift in the PCI-X development process, despite having useful expertise from creating the Accelerated Graphics Port (AGP) technology for plugging in graphics cards, Snyder said.
"Their input wasn't taken seriously by the technology working group," he said. "Intel had a lot of knowledge they learned from AGP. The result was a specification that required a lot of rework. That's what delayed PCI-X entering into the mainstream."
Tipley acknowledges PCI-X didn't arrive as smoothly as hoped.
"One thing we learned from that is we needed to bring it to the SIG a little bit earlier, and we're definitely doing that with Arapahoe," Tipley said. "There was a bit of electrical work that remained to be done after it was brought into the SIG. That's extended its definition a little longer than expected."
Or a lot longer. Compaq initially expected PCI-X to arrive in 1999, but it only began selling servers with PCI-X in recent months. Chips from ServerWorks, coming near the end of 2001, are expected to boost its mainstream use, while Intel plans its own support of PCI-X at that time or in early 2002, Snyder said.
"We're expecting the Arapahoe Working Group will have learned the lessons of AGP and PCI-X," Tipley said. "They'll just make new errors."