10-GigE Copper Heats Up
The 10GBase-CX4 standard, shepherded by the 802.3ak task force within the IEEE, was created as a low-cost alternative to fiber-based 10-Gbit/s Ethernet in the data center. But in the two years since CX4 got started, new advances have promised to make fiber modules cheaper. And a second copper standard, 10GBase-T, appears likely to uproot CX4 eventually.
None of this changes the fact that CX4 is available now and remains cheaper than any alternative. As a result, product announcements continue to emerge. Cisco Systems Inc. (Nasdaq: CSCO) included CX4 support in its latest raft of features for the Catalyst 6500. And Fujitsu Microelectronics America Inc. today added CX4 interfaces to its previously announced switch chip, which handles 12 ports of 10-Gbit/s Ethernet (see Cisco Beefs Up Catalyst, Fujitsu Enhances 10-GigE Switch Chip, and Fujitsu Packs in 10GigE ).
If CX4 is short lived, it will continue the streak of ill-fated copper standards for Ethernet. A similar scenario unfolded at the 1-Gbit/s generation, where a standard called 1000Base-CX was adopted, and it happened for 100-Mbit/s Ethernet as well. "You're going to see the same kind of evolution that's happened with Gigabit [Ethernet] over copper," says Steve Shalita, senior product marketing manager for Cisco's Gigabit Systems division.
The goal with CX4, which was ratified in January, was to let data centers avoid the cost of optics when dealing with short, rack-to-rack kinds of connections. "Removing the optics from 10-Gbit/s Ethernet is really a boon because you're bringing the cost down," says Asif Hazarika, product marketing manager at Fujitsu Microelectronics. In general, optical 10-Gbit/s Ethernet costs thousands of dollars per port, while copper interconnects could drop to a few hundred dollars.
But CX4 suffers on two fronts: distance and cabling.
The 802.3ak specifies distances of only 15 meters, and it's questionable whether that's enough even for connecting racks in the data center. "Nobody's really happy about this 15 meters. They'd like to have more," says Kamal Dalmia, technical marketing manager for Marvell Technology Group Ltd. (Nasdaq: MRVL).
Fujitsu's chip extends CX4 to 25 meters, and Mysticom Ltd. claims it's gotten CX4 to work at 30 meters.
But that might not be enough. The 10GBase-T standard, being developed by the 802.3an task force, calls for 50 meters and aspires to 100.
"It seems 50 to 60 meters is the maximum length for data centers," says Bruce Tolley, senior manager of emerging technologies at Cisco. "Fifteen meters doesn't get you very far in the data center."
Distance is key, but really, 10GBase-T is expected to win because it uses relatively common Category 5 copper cabling. CX4 was developed for InfiniBand cable, a more specialized variety that uses eight pairs of wire, twice as many as Cat 5. "When [10GBase-T] becomes available, there is no rationale for CX4 any more," Dalmia says.
Aside from copper concerns, CX4 faces a renewed challenge from the fiber side. In the time it's taken to finalize CX4, module vendors have created the XFP multisource agreement (MSA) for a serial 10-Gbit/s interconnect that's smaller and cheaper than previous alternatives. Separately, chip companies have banded together to create a standard for adding electronic dispersion compensation (EDC) to a 10-Gbit/s link, which would allow for the use of cheaper optics. The EDC option specifies 300-meter reaches on multimode fiber (see XFP No Longer a BFD and Vendors Still Driving LX4).
Copper is still cheaper, but these lower-priced alternatives, combined with fiber's ability to carry 10-Gbit/s signals farther, weaken the CX4 proposition, Dalmia says. "If the copper standard gets delayed too far out, it starts to lose its appeal. That's what happened to CX4."
Marvell introduced a CX4 chip in April 2003, and others -- including Broadcom Corp. (Nasdaq: BRCM) and Mysticom -- have followed suit (see Marvell Debuts 10-Gig Transceiver, Broadcom Tackles 10-Gig Copper, and Mysticom Demos 10-GigE Over Copper). Still, Marvell has no illusions about CX4's long-term future. "We're ratcheting down the emphasis on the CX4 side," Dalmia says.
Well, with all this doom and gloom, why support CX4 at all? For one, it's been completed as a standard, while 10GBase-T isn't expected to complete the IEEE ratification process until 2006. The 802.3an task force is still in its early stages; last week, the group heard technology proposals from companies including startups KeyEye Communications Inc., SolarFlare Communications Inc., and Teranetics Inc. SolarFlare, for one, expects to be shipping chips "by the end of the year," says Ron Cates, vice president of marketing.
Still, CX4 is in the driver's seat temporarily. "Since it's substantially cheaper than any of the other interconnects, and since data centers are the primary operation for 10-Gbit/s Ethernet today, we expect CX4 to be successful," Tolley says.
In Marvell's eyes, CX4 was a stage-setter, proving that 10-Gbit/s Ethernet could work on copper lines and -- more important -- putting pressure on optical vendors to bring prices down.
"If we didn't do CX4, the cost of optics would still be very high," Dalmia says. "CX4 has served its purpose from that perspective. It got the XFP group to move and got the EDC guys to move."
— Craig Matsumoto, Senior Editor, Light Reading