SAN MATEO, Calif. Top communications experts will debate how best to bring 10-Gigabit Ethernet to copper cabling at an IEEE meeting next month in Hawaii. Their efforts will serve as an important step in smoothing the evolution of tomorrow's business networks.
Industry players see the problem of how to run 10-Gbit/s Ethernet 100 meters over today's aging Category 5 unshielded twisted-pair cabling as one of the grand challenges of communications design. Indeed, many engineers say it can't be done.
But two startups SolarFlare Communications (Irvine, Calif.) and Plato Labs (Campbell, Calif.) are vying to show that it can despite huge hurdles in digital signal processing and analog design, as well as what some designers consider the physical limits of copper.
Those two newcomers face competition from entrenched players Broadcom Corp., Cicada Semiconductor Corp., Intel Corp. and Marvell Technology Group Ltd., each of which is developing its own solution. The strategies include upgrading to more-robust grades of copper cable, but all sides agree that no one knows which grade of copper cable will best accommodate 10-Gbit speeds.
The IEEE is billing its Nov. 10-15 meeting on Kauai as a call for interest on 10-GBase-CX4, a relatively quick, easy way to use Infiniband copper cabling to send 10-Gbit Ethernet signals over distances of less than 15 meters and to connect systems in a data center. Of perhaps more interest, though, will be SolarFlare's first open presentation on how to bring 10-Gbit Ethernet to Category 5 cables up to 100 meters, a tutorial organizers hope will catalyze interest in working out a standard for what would become 10-GBase-T. "This hasn't been openly discussed in public in the standards meetings before," said Chris DiMinico, chief technology officer with cable and optical-fiber maker CDT Corp. (Leominster, Mass.).
A 10-GBase-T standard would likely breathe new life into business networks. As part of a normal three-year upgrade cycle, companies are expected to refresh their desktop PCs starting mid-2003, the same time that Gigabit Ethernet becomes part of the standard PC configuration. Those systems should drive demand by mid-2004 for gigabit switches with 10-Gbit uplinks that will plug into the Cat. 5 infrastructure that makes up more than half of all the installed cabling in businesses today.
A 10-GBase-T standard won't spark an upturn itself, but it will push aside concerns about writing purchase orders for next year's gigabit-equipped desktops. "That's part of the premise. [The PC upgrade cycle] could be the beginning of a whole trickle-down effect in semiconductors and IT," said Jeremey Donovan, a chief analyst at Gartner Dataquest.
Why it's difficult
Currently the 100-MHz bandwidth of standard Cat. 5 wiring lacks the ability to handle 10-Gbit signals, said Sailesh Rao, a director of communications architectures in the former Level One group at Intel.
"If you do the math, you need to code about 50 bits over four pairs of wires, which is not possible because it implies about 4,000 to 5,000 levels of coding. You can't discriminate that many levels unless you go above the spec of Cat. 5 and then it's a whole new ball game," Rao said.
Vivek Telang, the chip architect who developed Cicada's 1-Gbit/s parts and is working on its 10-Gbit products, agreed. "You can make Cat. 5 work for 10 meters, but not for long distances over the existing cable plant. There's old cable that has been there for years and is in bad shape," he said.
Camden Ford, a senior product manager for Fibre Channel switch market leader Brocade Communications Systems Inc. (San Jose), was more blunt: "We've got to bite the bullet and start deploying fiber," he said.
SolarFlare and Plato Labs aren't revealing details of how they hope to crack the 10G-Ethernet-over-copper nut. But they do think they're on the right track. "It's a very, very hard problem that involves high-speed mixed-signal [technology], heavy comms theory and DSP. A lot of it is about the assumptions you live with about the capabilities of the channel," said George Zimmerman, founder, chief executive officer and chief technical officer of the 18-month-old company, now raising a second round of funding.
Zimmerman said he has gathered an unspecified number of DSL, Ethernet, network processor and wireless/analog experts to tackle the problem, which he says can be economically solved in 0.13-micron CMOS.
Zimmerman left his job as chief scientist at DSL specialist PairGain Technologies in 2000, when the company was sold to ADC Telecommunications, and began looking for startup opportunities to which he could apply his expertise in modulation techniques. An old friend from DSL circles, George Eisler, who also chaired the 1,000-Base-T standards committee, told Zimmerman the 10-GBase-T problem was the one to solve. A senior engineer at Sun Microsystems convinced him the market potential was huge.
While Zimmerman would not detail SolarFlare's work, he did offer insight into it: "If I had known how hard the problem was when we started I wouldn't have started it," he said.
Whether SolarFlare succeeds depends to a large extent on Zimmerman's ability to convince other top engineers that Cat. 5 offers previously undocumented capabilities that can support 10-Gbit technology, said Eisler, who is now a consultant on DSL issues for Mindspeed Technologies (Newport Beach, Calif.). Zimmerman's situation more or less parallels work Eisler oversaw on the 1-Gbit/s standard that produced the relatively recent Cat. 5e cable, which sports newly specified far-side interference and return-signal-loss properties.
Joseph Babanezhad, founder of Plato Labs, has had a separate team working on the 10-Gbit problem for two-and-a-half years, leveraging the company's six years as a developer of mixed-signal cores such as sigma-delta codecs and physical-layer (PHY) chips for wired and wireless nets. He claimed the company will design CMOS parts for 10-Gbit Ethernet over 100 meters of Cat. 5 for just two to three times the cost and power consumption of today's 1-Gbit/s devices. The parts should also see use in future Infiniband, Fibre Channel and iSCSI systems, Babanezhad added.
"I am convinced this thing will happen. Just on a technology basis, it could happen in less than a year," Babanezhad said. "There's a change of paradigm that needs to happen here, though. In terms of implementation, new things need to be done. There's no question that you have to push Cat. 5 to higher frequencies."
Perhaps just as tricky as finding a solution is codifying it in a standard. "It's hard to point out any successful Ethernet product that was not based on a standard," said Nariman Yousefi, senior director for engineering development in the net group at Broadcom.
But standards committees often devolve into "heated battles to prevent any one camp from defining the line codes and modulation scheme," said Cicada's Telang, who claims he made the first public presentation of 10-Gbit Ethernet over UTP way back in January 2000.
"It was too early. Now that 1G is getting established, we need to look at the next level," Telang said.
For newbies like Solar and Plato, "this is their big play and they have to make a splash. We'd be more cautious now," Telang added.
Perhaps the biggest battle this time around will be over cabling. Intel's Rao said his group will propose next year a move to newer and more expensive 600-MHz Category 7 cables. He is in the process of determining what symbol rates and coding schemes to use for it. "We know it's feasible, we're just working out the details," he said
"If we don't get push-back from end users, I'd use it in a heartbeat because it makes for more simple silicon," Telang noted, "but my view is Cat. 6 is a realistic compromise solution. To get to any viable distance you will need Cat. 6." Specifically, Telang said, four-pair Category 6 cables are best for 25-meter stretches of 10-Gbit Ethernet.
Cicada has already developed "pieces" of such a solution, including high-frequency phase-locked loops, high-precision A/D converters and signal processors, Telang said.
Many chip designers forget thorny problems like alien crosstalk from outside the cable, said CDT's DiMinico.
"You need something like Cat. 6 to get over this. I don't think they should go for Cat. 5. [Advocates] may make themselves famous, but it will be in 2010. I'd rather see them start with Cat. 6," he said.
Broadcom's Yousefi said a more sensible approach to switching cable is to stick with Cat. 5 but reduce the maximum length from 100 to about 50 meters, to reduce silicon costs and power requirements. Switch and server makers looking to link their systems would be the biggest users and would be served by such lengths, he said.
"One hundred meters is a stretch goal. You can stick with that standard cable length by adding silicon complexity, but how many people will vote for that?" Yousefi asked.
Broadcom will also push for a standard that allows for 1- and 10-Gbit products like today's 10- and 100-Mbit chips used to bridge the technology transition.
Kamal Dalmia, a marketing manager for PHY chips at Marvell, said his company will support sticking with Cat. 5. "It's extremely challenging, to say the least. We're actually doing early feasibility studies on that," he said.
"Once you go to Cat. 7 you've lost the operational advantage," Zimmerman said. Others note that Category 7 cables provide better bandwidth and shielding but are heavier, more costly and less flexible than Category 5 and unsuitable for installation over hung ceilings, where Cat. 5 is used today.
"Cat. 7 is out of the question. That would be a difficult installation problem," Eisler said. "There's a reasonable argument you could make for going with Cat. 6 but stretching it to Cat. 7 would make no sense."
50 meters and counting
For its part, Mysticom (Mountain View, Calif.) believes it can push 10-Gbit Ethernet up to about 50 meters over copper. However, the commercial goal for the company's next-generation transceiver is for longer optical fiber stretches rather than copper ones.
"We don't see any customer interest in 10G Ethernet over 100 meters of Category 5 wire, but we do have interest in transceivers for longer fiber and Infiniband cable links," said Amir Bar-Niv, director of engineering at Mysticom.
Mysticom's next-generation transceiver, which has already taped out, will use a proprietary encoding scheme that will enable signaling to extend 600 to 800 meters over low-grade fiber.
One router maker is interested in that capability as part of a 10-Gbit Ethernet backplane, and a switch maker wants to use the chip for fiber links between distributed switches in a data center that might stretch between multiple buildings.
Bar-Niv said the new transceiver might allow 10-Gbit Ethernet signaling to extend about 20 meters over Cat. 5 cables.