I/O Moves Into The Express Lane

As PC microprocessors pass the 3-GHz clock-speed mark and Gigabit Ethernet begins its move onto the corporate desktop, it's easy to forget that the typical expansion bus in desktops sold today still runs on 11-year-old technology.

The 32-bit, 33-MHz Peripheral Component Interconnect (PCI) bus specification, which debuted in 1992, remains the standard. PCI supports an aggregate bandwidth of just 133MB/sec. (1Gbit/sec.), and that bandwidth must be shared among all devices on its multidrop bus. "I/O has moved at an anemic pace," says Tom Bradicich, chief technology officer for IBM's xSeries server line.

An emerging I/O standard, PCI Express, is about to speed things up. The current specification, supported by the PCI-SIG vendor consortium and promoted heavily by Intel Corp., succeeds PCI's parallel bus design with a high-speed, point-to-point serial interconnect technology that Intel and others say will be both faster and more reliable. It will deliver an aggregate bandwidth ranging from 5Gbit/sec. (500MB/sec.) to 160Gbit/sec. (16GB/sec.). And because it's serial, the technology will work more efficiently with other emerging high-speed serial technologies such as Serial ATA and InfiniBand. It will also offer hot-plug and hot-swap support and enable direct peer-to-peer communication between attached devices without involving the processor chip set.

But PCI-SIG isn't just promoting PCI Express as a desktop PCI expansion bus replacement. The consortium is positioning it as a general-purpose I/O technology for desktops, workstations and servers that will also speed internal chip-to-chip and graphics adapters. Intel plans to release a PCI Express chip set for servers, code-named Lindenhurst, early next year, with a PC chip set to quickly follow.

On the desktop, the first PCI Express add-in devices are likely to be graphics chips because the Accelerated Graphics Port (AGP) standard has run out of gas, says Jim Pappas, director of technology initiatives for Intel's enterprise platform group. "All new graphics development [beyond AGP 8X] is being done on PCI Express," he says, noting that the first graphics products could be available by early next year.

Although vendors generally view a migration to serial I/O technology as inevitable, PCI Express' near-term prospects beyond high-end graphics are less clear. IT managers say desktops don't yet need the extra I/O bandwidth. "People are not taxing the equipment they have," says Scott Newton, director of IT at Otis Spunkmeyer Inc. in San Leandro, Calif.

PCI Express also represents a fundamental technology shift in that its physical interfaces, which include desktop expansion slots, a new PC Card format (code-named Newcard) and Mini PCI Express slots, won't accept existing PCI devices.

In the server world, where I/O bandwidth is more of an issue, PCI Express faces competition from enhanced versions of the PCI bus standard, including the emerging, 2GB/sec. PCI-X 2.0, due out later this year. But vendors are split on whether to stay on the PCI-X road for backward compatibility or go directly to PCI Express.

"We are focused on PCI-X 2.0 in the 2004-2005 time frame. It provides improvements in performance and backward compatibility for investment protection," says David Heisey, manager of advanced technology at Hewlett-Packard Co. IBM's Bradicich agrees that the current PCI specification "does not comprehend the requirements for servers in the enterprise."

But Jimmy Pike, director of server architecture at Dell Computer Corp., says his company plans to pass on PCI-X 2.0. "Everybody is already admitting that there is some reasonably short life to the stuff they're putting on PCI-X 2.0. Your investment is better protected with PCI Express," he says.

The issue may be decided by chip-set makers, since the use of PCI Express internally makes deploying the technology for external I/O devices easier.

"Whoever is providing the chip sets will determine whether PCI Express will appear in systems," predicts Bradicich. Intel dominates the desktop market. Santa Clara, Calif.-based ServerWorks Inc. has the lion's share of the server market, and it plans to focus on PCI-X 2.0 chip sets, says Virkam Karvat, director of marketing at the Broadcom Corp. subsidiary. He doesn't expect to see PCI Express in servers until 2005, which is when Bradicich says IBM plans to include it in the xSeries line.

Even then, PCI isn't likely to go away anytime soon, which means IT will be managing yet another I/O interface.

Motherboard of Invention

While servers have adopted faster versions of PCI, most desktop vendors have stayed with the original standard because, outside of graphics, desktops haven't been I/O bound.

Once users need the bandwidth, PCI Express could take off because Intel's dominance in PC chip sets gives it the power to push PCI Express into the majority of desktops, says Bert McComas, an analyst at InQuest Market Research in Higley, Ariz. "Since it's on the motherboard, [vendors will] wire it up to a slot or two or three," he says.

Chuck Stancil, a technical staff member in HP's personal systems group, agrees that desktops will eventually go to PCI Express. The question is when.

PCI Express will gain a foothold in 2004 by replacing the AGP graphics slot, enabling faster high-end graphics and streaming video applications on desktops. But its use for general-purpose I/O could be much further out.

Nathan Brookwood, an analyst at Insight64 in Saratoga, Calif., predicts that other high-performance I/O applications will come once systems are equipped to take advantage of them. And one application is already here. "Gigabit Ethernet really can saturate a desktop PCI bus. To get around that, vendors are doing all sorts of funny, ad hoc solutions," Brookwood says.

Tony Pierce, chairman of PCI-SIG, says he expects Gigabit Ethernet will become standard in desktop systems within 12 to 18 months, which could give PCI Express some momentum.

The technology will likely phase in slowly, though. "There isn't a compelling reason to ditch PCI slots, which are dirt cheap," and replace them with PCI Express, says Stancil. And though vendors say PCI Express ultimately will be less expensive to deploy, the need to support both PCI and PCI Express means that early hybrid machines are likely to be more costly. Without a compelling need, the complexity, compatibility and cost issues could leave many organizations on the sidelines.



Where PCI Express Fits

PCI Express, designed as a general-purpose I/O technology, could replace the proprietary “mezzanine bus” communication technologies used for internal, chip-to-chip communications, as well as AGP for graphics I/O and PCI for expansion cards. For hard disk drive I/O, ATA technology will migrate to its own serial standard, called Serial ATA, and will access the I/O hub in the chip set directly \[see “Serial ATA Takes on SCSI.”

Where PCI Express Fits
Source: Intel Corp.

Copyright © 2003 IDG Communications, Inc.

7 inconvenient truths about the hybrid work trend
Shop Tech Products at Amazon