Global Sources
EE Times-Asia
Stay in touch with EE Times Asia
EE Times-Asia > Networks

A link forged in copper

Posted: 16 Mar 2005 ?? ?Print Version ?Bookmark and Share

Keywords:copper? keyeye communications? solarflare communications? ethernet? broadcom?

Two design teams that helped define high-bit-rate digital subscriber line standards a decade ago are facing off in yet another realm where copper carries data. The founders of KeyEye Communications Inc. and SolarFlare Communications Inc. are again confronting line-coding and system noise problems as they work on conditioning copper wire to ferry 10Gbps Ethernet frames over short distances in the data center.

But the quest could prove more vexing than the mid-'90s work to groom unshielded twisted-pair for the wide-area network. Indeed, KeyEye and SolarFlare may end up being the only 10GbE startups capable of sustaining a physical-layer interface IC business against such giants as Broadcom and marvell. Key figures in that drive will be their respective CTOs, both critical players in the earlier HDSL effort, who today are collaborators under the IEEE's 802.3an task force as well as competitors in the field.

In the mid-1990s, Hiroshi Takatori was the chief architect for HDSL-2 chips at Level One Communications Inc., a Sacramento, Calif., analog chip company later acquired by Intel Corp. George Zimmerman was chief scientist at PairGain Technologies Inc., one of the first OEMs to develop access equipment for first-generation HDSL networks.

Level One and PairGain were separately promoting pulse-amplitude modulation coding schemes for HDSL-2 that faced tough competition from the "partially overlapped echo transmission," or Poet, algorithms favored by Siemens and Adtran. Takatori and Zimmerman started out as opponents in their means of implementing PAM, but over time Takatori moved toward support of the "overlapped PAM transmission with inter-locking spectra," or Optis, code developed by Zimmerman. With prodding from Level One, PairGain and ADC Telecommunications, which later bought PairGain, Optis became a standard for HDSL-2.

Following Intel's 1999 acquisition of Level One, Takatori became an Intel Fellow working on physical-layer networking problems. Stephen (Mike) McConnell, who had coordinated data-communications planning at Level One, took a similar post in product planning at Intel.

"He [Takatori] had been aware of common coding issues involved in DSL and Ethernet, and he was prepared to be co-founder when I developed the business plan to create KeyEye in late 2001," said McConnell, who today is VP of strategic marketing at the Sacramento company.

Zimmerman, meanwhile, had been looking for new challenges in signal processing since ADC picked up PairGain in mid-2000, and the plans for SolarFlare evolved along a similar timetable. Before PairGain, Zimmerman had headed signal processing for NASA's Deep Space Network at the Jet Propulsion Laboratory. His primary interest has remained practical coding for carrying data at high speeds and over long distances.

When the IEEE 802.3 working group first entertained concepts for 10Gbps networks early in this decade, the assumption was that short-reach fiber, long-reach fiber and coaxial cable were the only feasible options as physical media. It took the continued prodding of KeyEye and SolarFlare founders to convince skeptics to create a 10GBase-T study group for 10Gbps Ethernet over copper, later dubbed the 802.3an task force.

Proving the market for copper-interface 10GbE chips, however, may be as difficult as implementing signal-processing algorithms in silicon. That's as much a reflection of the changing nature of high-speed Ethernet as it is the tough postrecession business climate. In the WAN, 10Gbps backbones have become common, in part because of the indefinite postponement of 40Gb upgrades. In the enterprise LAN, trends aren't so simple.

Every generation of Ethernet before the current one has been driven by a "race to the desktop." This time around, bringing multiple gigabits per second to the desktop takes a backseat to using Ethernet as a high-speed serial connection for linking servers, storage clusters and switches in the data center. And at 10Gbps, the boutique applications in the data center may make high-speed copper implementations more problematic.

A host of fiber options are competing with 802.3an. But because 10GbE is seen as a pure data-center play, where short links would provide Ethernet framing among Web server clusters, the first generation of products would not need a star-based switched-packet topology. Ethernet in such an environment would be used as a serial alternative to Infiniband or Myrinet. Since the distances are shorter, signal conditioning demands are not as great and opportunities for copper thus appear feasible.

In the first months following 10GBase-T's formation, SolarFlare demonstrated implementations of a 10Gb network using a special grade of Category 5 wiring. Over the course of several IEEE meetings, however, participants decided that links of up to 55m would require a higher-grade cabling, called Category 6, while links extending to 100m would require either a special subclass of Category 6, called Class E, or a cabling used in Europe called Category 7.

Is this asking for too great an infrastructural upgrade? The key, said KeyEye's McConnell, is to make a convincing case for a mixed bag of short-reach fiber, coaxial cable and Category 6 or 7 wiring, each for different interconnect tasks.

This cabling problem is precisely what concerns many involved in the early 10GBase-T meetings who have since pushed out their development schedules. Broadcom Corp., Vativ Technologies Inc. and Marvell Semiconductor Inc. were strong voices in those early meetings, but their interest in designing ICs to meet the target standard, slated for completion in early 2006, has cooled.

"In general, we're very negative on the success of 10GBase-T, because the data-center market cannot drive integration of first-generation designs," said Sreen Raghavan, CEO of Vativ Technologies. Vativ has active development programs in several 10-Gbit fiber standards, "but we see several aspects of 10GBase-T as ridiculously expensive for the applications considered," Raghavan said.

Zimmerman of SolarFlare said he expects some larger semiconductor players that must live by high volumes to wait for a market to mature in 2006, but he said that SolarFlare founders remain convinced that "10GBase-T is a market with a potential in the double-digit millions."

KeyEye, meanwhile, has not been dissuaded by concerns over market size. Early on, the startup considered focusing on multigigabit copper backplanes, but it has since broadened its horizons to consider any interface that can carry data at 10Gbps over copper. The company has established a patent portfolio in a wide range of signal-processing algorithms. Its EchoWave software suite, to be implemented in CMOS products for multigigabit designs, handles interference, mismatches and crosstalk. "We have taken a heavily analog signal-processing approach and are aiming for achieving the lowest possible power dissipation and smallest die," McConnell said.

The goal is not simply to reduce the footprint for line cards that use IEEE-standard 10GBase-T chips with direct copper interfaces. McConnell said a core KeyEye strategy is to embed copper-interface chips in an optical package adopted as a multisource agreement, "achieving IEEE compliance through the MSA cage." Since an optical package requires transmit and receive optical subassemblies, along with silicon and photodiode interfaces, an MSA using a copper interconnect will always show cost advantages over an optical solution, McConnell said. Yet the copper solution can use a line card originally designed for fiber 10Gb links.

The problem in designing for the limited-socket data-center market is that startups struggle to gain the revenue stream necessary to perfect 10GBase-T designs. Indeed, young companies focused solely on 10GBase-T have struggled. SolarFlare had to juggle some staff in 2004 to save money. Teranetics Inc. has limited its staff to a core design team. And some working-group players like Plato Labs Inc. have yet to grow beyond angel-investment size.

The IEEE 802.3an task force has completed Draft 1.1 of a 10GBase-T standard and is aiming to have Draft 2.0 completed at a March meeting in Atlanta. The task force has agreed on a Tomlinson-Harashima precoding method for pre-equalization before transmit and on low-density parity check codes within forward error-correction blocks.

- Loring Wirbel

EE Times

Article Comments - A link forged in copper
*? You can enter [0] more charecters.
*Verify code:


Visit Asia Webinars to learn about the latest in technology and get practical design tips.

Back to Top