networkZONE Products for the week of March 29, 2004
SolarFlare Says . . .
10-Gbit/s Breakthrough! -- SolarFlare Demonstrates World's First 10-Gbit/s Ethernet Transceiver Running On Cat5e Copper Cable
Promotes the Development of the 10GBase-T Ethernet Standard Currently Underway
SolarFlare Communications Inc. has created a demonstrable, complementary metal-oxide semiconductor (CMOS)-based transceiver that is capable of sustained 10 gigabits-per-second (Gbits/s) operation over industry-standard Category 5e (Cat5e) copper cable. Cat5e is the cable type most frequently used to interconnect computers, switches, and routers in Ethernet networks.
"Today's 10Gbits/s Ethernet links are implemented with either very expensive optical-fiber transceivers or with cumbersome, short-reach, twinax coaxial cables," said Ron Cates, SolarFlare's vice president of marketing. "The advent of a transceiver capable of operation over the installed base of Cat5e cable defies industry 'experts' who claimed it couldn't be done and promises to dramatically lower the costs of 10Gbits/s Ethernet deployment and increase its utilization in enterprise computing networks and data centers."
SolarFlare's transceiver uses industry-standard CMOS fabrication technology and proprietary signal-processing algorithms that improve data recovery and mitigate noise in copper wires so that they can support higher data rates. The transceiver is designed to operate over Cat5e, Cat6, or Cat7 cables with lengths up to 100 meters.
"While several companies have speculated about the possibility, and even showed feasibility through simulations, I am incredibly proud to say the SolarFlare team is the first to demonstrate 10Gbits/s operation over long distances of copper cable with real hardware," said Russell Stern, SolarFlare's CEO. "This achievement marks a major milestone and will significantly boost the initiatives that SolarFlare and other Ethernet industry leaders are taking in formulating an industry standard for 10 Gigabit Ethernet transmission over twisted-pair copper cable under the auspices of the Institute of Electrical and Electronics Engineers (IEEE) 802.3 committee."
"SolarFlare's ability to implement 10Gbits/s-over-twisted-pair cabling represents the fourth generation of Ethernet transceivers that have historically catalyzed the adoption of a faster version of Ethernet by removing the need to deploy expensive optical fiber infrastructure in enterprise networks," said Greg McAdoo, formerly with Cisco Systems and now a partner with Sequoia Capital -- one of SolarFlare's lead investors.
The SolarFlare 10Gbits/s Transceiver digital signal processor (DSP) is a mixed analog and digital single-chip device fabricated with 0.13-micron CMOS technology and housed in a 600-pin ball grid array (BGA) package. Designed in-house and containing in excess of four million logic gates, the transceiver was designed to enable high-volume production in a customer-owned tooling (COT) flow. The device is capable of direct interface to existing 10Gigabit Ethernet media access controllers (MACs) and uses a standard management data input/output (MDIO) control bus.
EN-Genius Says . . .
I normally refrain from reviewing products ahead of their commercial release, but occasionally make exceptions, such as in the case of SolarFlare. If their assertions about the cost and performance of their recently announced pre-standards 10-Gbit/s copper Ethernet PHY are true, the high frontier has moved back from the backplane world to the network. Definition of 10G Ethernet-over-Cat-5e standard is now on the fast track -- in no small part thanks to SolarFlare's roll-out of a working technology demonstration which helped cut through skepticism and inertia. While there are some significant technical hurdles to overcome (not to mention the grueling politics of the IEEE standards committee), SolarFlare seems to have at least given us a hint of what's possible.
I'll try to provide brief look at the basics of the chip's innards, and give you an idea of why I think these folks are on the right track. For a detailed explanation of the technology behind the chip, SolarFlare has kindly supplied us with an excellent white paper that should make for some interesting reading. You can click here to access the white paper which includes the block diagram of the part.
A Peek Inside
Much like its 1-Gbit/s counterparts, the prototype device uses the four twisted pairs of cable as separate data pipes, each carrying a quarter of the overall stream -- in this case 2.5 Gbit/s, plus coding overhead. Because of the inherent bandwidth limitations of Cat-5e cable, the prototype chip encodes each of the four 2.5 Gbit/s channels in a 3 bits/baud scheme using PAM-10 coding. The circuit optimizes receiver BER performance by using a 4-dimensional trellis-coded Viterbi algorithm for symbol detection. It's somewhat similar to the scheme used in 1000BASE-T and provides sufficient SNR improvement to provide a theoretical BER level of 10-12.
Even at the lower frequency range that PAM makes possible, the copper cable's attenuation characteristics play hell with the signal and requires some pretty sophisticated equalization to provide the amplitude resolution needed to accurately fish the PAM-10 signal levels out of the mud. Mr. Cates was not very forthcoming about the details of the EQ scheme, and neither is the paper, but I'm inferring it's an adaptive scheme, most likely using decision feedback of some sort.
The big questions in my mind are how many taps they are running, how much of the EQ is being done in the digital domain, and whether they are equalizing on a per-symbol basis. At the risk of making an idiot of myself, I'll infer a few things from the block diagram they supplied. For one thing, they have wisely chosen to have an external AFE for the development chip set which houses a PGA and ADC stage. This indicates to me that the PGA does some rough EQ as it levels out the signal, but the bulk of it is performed in the digital domain within the main chip. Too bad the diagram they supplied does not have more detail on the EQ, but that's where a manufacturer can put some really juicy "secret sauce" -- even in a standards-based product. The press release explains that the larger DSP is fabricated in 0.13 micron CMOS but is mum about how they build the analog front end (AFE). Naturally, this makes me wonder if the AFE is also fabricated in "vanilla CMOS" or in some bipolar technology that provides better linearity and noise characteristics. If it were a non-CMOS part, SiGe would be my bet.
The other critical area of a transceiver is the cancellation for NEXT, FEXT, and a new category that SolarFlare refers to as "alien," or signals originating outside the cable from non-network sources. From what I am able to infer from the white paper and conversations from Mr. Cates, their hybrid approach to cancellation is reasonable because it reduces the amount of FIR-based DSP to a level that might fit on a commercially-viable semiconductor. Since SolarFlare has not shared any of the details of how they accomplish such a Herculean task, I'll take an optimistically-skeptical stance and concede that they've proven that sending 10 Gbit/s over Cat5e is possible, but am undecided as to whether they've come up with the right approach.
SolarFlare was very clear that they are running pre-standards silicon -- for technology demonstration purposes only -- and that release of their production version is scheduled to coincide with the appearance of the draft standard, probably around the end of the year. The architecture of the prototype is highly programmable and will probably track "most" of the changes as the standard evolves. This leads me to expect a smaller, less-programmable production version of the DSP chip.
The ability to move 10 Gbit/s over Cat-5e/Cat-6 cabling would fill a large natural demand since fiber is still expensive and 1G switches are becoming commodities and need 10x uplinks. SolarFlare's VP of marketing, Ron Cates, also pointed out another potential source of "market pull" that I'd missed -- he perceives an earlier big market for high-speed pipes for linking clustered computers that support enterprise applications in a scalable manner. Of course, there should also be a big potential market in storage networks since 10G copper Ethernet will likely be less expensive than similar FibreChannel solutions.
Deployment costs also favor the technology because its ability to work with widely-accepted Cat-5e and Cat-6 cabling lets it work with much of the installed base of cabling, and it can be put in with little or no additional training of installers. I pointed out that CX-4 is currently an attractive option, but Cates pointed out that it is still more expensive than the projected cost of 10G Ethernet -- not to mention that the cost and bulkiness of the Infiniband-style cabling are limiting acceptance.
Of course this natural market need does not always translate to a natural solution -- especially when it's this close to the bleeding edge of technology and basic physics. For this reason I look forward to seeing the demonstration they'll be running at Electronica-USA 2004 this week as I'll be able to learn a bit more about how the technology works in it's pre-standards form. But since technology demonstrations are under rather controlled conditions and can cover up any number of unresolved issues, I'll hold my final judgment until I know more about the chip set's architecture.
|
