CNET también está disponible en español.

Ir a español

Don't show this again

Water on the moon Microsoft Surface Duo Stimulus negotiators MagSafe accessories iPhone 12 and 12 Pro review Obi-Wan Kenobi Disney Plus show Murder hornet nest

Intel-server debate a pain for industry

The chip giant and ServerWorks, the chief maker of chipsets for Intel-based servers, can't agree on an important design issue--and that's bad news for HP, Dell and other server makers.

SAN JOSE, Calif.--Intel and the chief manufacturer of chipsets for Intel-based servers don't see eye-to-eye on an important design issue--and that could create problems for Hewlett-Packard, Dell and other server makers.

Complete coverage
Intel Developer Forum news
Read what visionaries at the confab
expect for wireless, security

The debate between the chipmaking giant and Santa Clara, Calif.-based ServerWorks involves which of two technologies should become the standard for plugging devices such as network adapter cards and storage system controllers into future servers. Intel stumped for the technology it favors at the Intel Developer Forum here this week.

Servers are networked computers that handle important tasks like routing e-mail. Their central processors are linked to other components via an essential device called a chipset. If the processor is the brain of a computer, the chipset is the spinal cord, connecting to everything else. In this case, the chipset serves as a link to the slots for network adapters and other cards.

Agreement between Intel and ServerWorks on one standard technology for dictating how these slots work would simplify life for makers of Intel-based servers and for companies that make the cards that plug into the slots. It would free them from having to choose between the technologies themselves or from having to build machines that support both.

If forced to make a choice, they run the risk of picking one technology that ultimately loses out to the other, in much the same way the Betamax format lost out to VHS in the video market. And their customers typically use servers for three years or more and don't want to get locked into technology that ultimately falls by the wayside.

The technology Intel favors is called PCI Express, formerly called 3GIO. Intel developed PCI Express for desktops and laptops initially but believes it has a place in servers. ServerWorks, however, is likely to all but veto that plan.

It prefers PCI-X 2.0, an update to the PCI technology. Intel argues PCI Express will be less expensive, but ServerWorks counters that PCI-X 2 provides for a smoother transition from current technology.

Intel has powerful PCI Express allies, notably No. 2 Intel server seller Dell Computer, which joined Intel to trumpet PCI Express during the Intel Developer Forum this week. But outside experts think ServerWorks will prevail initially.

"It looks like out of the gate, ServerWorks is going to lead the next round in terms of driving standards toward PCI-X 2.0," UBS Warburg securities analyst Alex Gauna said.

The company, a subsidiary of communications chipmaker Broadcom, is no lightweight. ServerWorks specializes in chipsets and is the dominant supplier for servers built around Intel processors. It has about 90 percent of the Intel-server chipset market, Gauna said, both in terms of revenue and unit shipments, and its components are used in most of the Intel servers sold by Hewlett-Packard, Dell and IBM. Intel itself also makes chipsets, but its percentage of the market is comparatively tiny.

Intel will use PCI Express in all its chipset designs beginning in 2004, said Jim Pappas, director of technology initiatives for Intel's Enterprise Platforms Group.

"PCI Express is extremely important in servers," Pappas said. It uses a better foundation technology--"serial" connections rather than the "parallel" data pathways used by PCI, PCI-X and PCI-X 2.0--that everyone eventually will have to adopt. "We believe it's an inevitable transition. The more rapidly we address it, the more competitive platforms we're going to have, with better cost and better performance."

A third strike for Intel?
But Intel has some black marks on its record when it comes to influencing computer designs beyond the processor. Years of backing Rambus memory technology came to nothing. And while Compaq Computer, IBM and HP successfully pushed PCI-X--the first extension to PCI--Intel advocated a PCI Express predecessor called Next-Generation Input-Output. NGIO became part of the InfiniBand technology that Intel eventually dropped.

"After racking up one strike with Rambus and another with InfiniBand, the last thing Intel needs is a third strike with 3GIO," said InQuest Market Research analyst Bert McComas in a report this month. "3GIO is comparatively late, offers little in the way of performance benefits and introduces new complexities in silicon integration and motherboard design."

Kimball Brown, vice president of business development at ServerWorks, said future versions of PCI Express hold potential, but today's technology just isn't worth it. It will take 16 PCI Express links bundled together to reach the performance of the faster version of PCI-X 2, called 533.

Intel disagrees. Pappas said eight PCI Express links will beat out PCI-X 2 533 because PCI Express can simultaneously transfer data in two directions, while PCI is one way.

Dell buys Intel's argument, chiefly on the basis of cost, said Joe Sekel, senior server architect at Dell. PCI Express is plenty fast, he said. It's also geared not only for plug-in cards but also for high-speed networking chips soldered straight into the computer. And sooner or later companies need to switch to serial communications.

"At this point, our plans are to go for PCI Express out of the chute," Sekel said, though he didn't rule out PCI-X 2.0 support, given issues around customer preferences and the ever-present possibilities for schedule slips.

The technology issue of serial vs. parallel communications boils down to the best way to get information from one point to another. Parallel communication uses several wires--64 in the case of PCI-X 2.0--that carry signals that travel in lockstep. But it's difficult to make sure the lines are synchronized because the wires typically extend different distances as they turn corners and suffer different interference depending on what other signals are nearby.

Rather than forcing data transfers to march to the beat of a single clock, serial communication sends information, including timing information, down high-speed lines that aren't synchronized with one another. A "deserializer" at the far end of the lines has the job of pulling out sequence and timing information from the high-speed lines and reconstructing the original data.

In the long run, it seems likely serial connections will win out.

ServerWorks' Brown believes the high-speed serial technology in PCI Express has merit eventually--for example, later in the decade, when each line can transfer data at 5 gigabits per second instead of the 2.5 gigabits per second of the first versions of PCI Express.

The important factor is whether to make the change now or later.

"Change is always difficult," Dell's Sekel said. "What we look for is the right we only have to do it once."