i/oZONE Products for the week of October 27, 2003
Quellan Says . . .
Gigabit Backplane Breakthrough? -- Part 2 of a 2-part
review of Quellan's Gigabit Noise Canceller
Live demonstration shows BER performance improvement by a
factor of 1,000,000 in conventional backplanes is possible at 5 Gbit/s speeds
Quellan Inc, a leading developer of Gigahertz spectrum Analog Chips, demonstrated the ability to "cancel" unwanted "crosstalk" noise in a multi-Gigabit backplane at the IMAPS Conference. This cancellation technology enables existing noise impaired equalized backplanes to go from passing 100,000 bits before an error occurred to a throughput of 10 trillion bits before an error occurred -- a Bit Error Rate improvement of 1,000,000. This translates into a throughput improvement from 1.25Gbps to 7Gbps while sustaining a BER of 10e-17.
Similar in principle to the technology used in consumer "Noise Canceling Headphones", Quellan's technology cancels noise at frequencies 1,000,000 times higher - in the Gigahertz range. As system and network traffic bandwidth increases, the aging backplanes in the field- those passive chassis that carry signals between cards - become increasingly unable to transport signals from card to card. The noise that is coupled from pin to pin on the connectors increases with frequency and at the rates required, render equipment inoperable. Since this equipment has been installed for years, it cannot be torn down to be upgraded. Quellan's chips allow the infrastructure to stay in place and by incorporating chips on new switch cards, speeds of 5 and 6.25 Gigabits per second can be achieved on legacy backplanes. IT managers benefit by leaving their network intact and OEMs benefit by being able to offer higher performance capabilities into existing systems, which greatly reduces development cost and time to market.
"Crosstalk is now a significant barrier to upgrading all types of network, compute and broadcast equipment" said Tony Stelliga, CEO of Quellan. "In fact, we have seen systems where crosstalk at desired data rates exceeds the signal level. By canceling the undesirable 'aggressor' signals, we are able to allow the 'victim' signal to propagate freely down the backplane, largely unaffected. The use of our chips on new line cards, then, allows in-field system upgrades in the compute, telecom and enterprise and broadcast markets."
"Quellan's demonstration today clearly showcases a valuable technology to the communications marketplace," said Mike Noonen, GM of the Network Technology Group at National Semiconductor. "Quellan has a approached this imminent problem from the system level and have developed a winning technology to address it."
Quellan ran the demonstration on their CSP Real Time or "CSP R/T"
system that is now available to OEMs to perform evaluation of Quellan's
Technology and to better characterize their backplanes. The CSP R/T provides
single and dual channel cancellation and a Feed Forward Equalizer for high
performance cancellation and system analysis using a range of Pods that
are available for different connector sets.
analogZONE Says . . .
It's standard practice in this business for a tech writer to evaluate chips, software, and other products that you have not actually seen operate, or may not even exist yet. We can usually do this in a semi-intelligent manner because most of the things we write about are extensions of existing technologies and employ commonly understood principles of operation. But on those occasions when you get a so-called disruptive product, it's difficult to make a good analysis of the manufacturer's claims without doing a bit more homework than usual.
A good case in point is Quellan's Nx600 hybrid-architecture signal canceller, a device designed to increase the available bandwidth in the backplanes of high-capacity routers, switches, and other networking equipment. Quellan had made some significant claims for their new chip that sounded like science fiction the first time I heard them. Their assertions that their canceller could achieve a six-order-of magnitude reduction in cross-talk seemed to be about as plausible as the fictitious "Oscillation Overthruster" device featured in the cult sci-fi movie Buckaroo Banzai which supposedly allowed vehicles to drive through solid matter.
analogZONE's Editor-in-Chief, Paul McGoldrick, was less skeptical than I was. Paul attended my joint briefing with Quellan and rightly pointed out that the cancellation techniques employed by the Nx600 have been used in radar signal processing and elsewhere for at least 20 years. He also saw a number of interesting applications for their product that I did not catch. You can view Paul's comments by clicking here .
Despite Paul's sanguine approval, I still had my doubts. For one thing, it appears that Quellan's canceller chip is the first time those principles have been implemented in such a highly integrated manner -- at least in a civilian application.
After some discussion with Tony "John" Stelliga, Quellan's CEO, it was agreed that I'd get some hands-on time with their product before drawing any conclusions. What I got to see was a demonstration using pre-production prototype chips that are far less integrated than the Nx600 will be when it comes out of the fab in a few weeks. Nevertheless, results they produced were enough to convince me that Quellan's approach has the potential to solve a wide class of technical problems beyond simple backplane application, including wireless networking, cellular phones, and digital video distribution in broadcast studio equipment.
A Brief Recap
For those of you who did not catch the first part
of this review, Quellan's chip adds cancellation to the pre-emphasis and
equalization techniques commonly used by designers to drive higher bit rates
across the copper pairs that make up a backplane. As Quellan CTO, Dr. Joy
"Monkey-Boy" Laskar, points out, a certain percentage of backplane
channels have enough cross-coupling to adjacent traces that pre-emphasis
only adds to the cross-talk, while equalization simply brings up the noise
floor along with the signal itself. Clearly what's needed here is a cancellation
mechanism that samples the noise signal on the adjacent channel and subtracts
it from the signal, but that's a tall order when you are working with data
streams running at 5 Gbit/s or more.
Digital cancellation techniques have been commonly used in lower-speed applications, such as DSL, digital cable, and 100/1000Base Ethernet. Broadcom, TI, and others have made a good part of their fortunes making CMOS-based DSP engines that are very effective at reducing the near-end cross talk (NEXT) caused by reflections at connectors (and other impedance disruptions), and coupling cross-talk (CoXT) caused by capacitive coupling between adjacent wires. Quellan rightly contends that all-digital techniques wind up being impractical (power-hungry and expensive) at speeds of a Gigabit and up -- mostly because of the speed limitations of today's CMOS processes. To get around this, they developed a chip that performs most of the actual signal processing functions in the analog domain, but uses digital logic to configure and control the process. Accelerant has applied a similar hybrid approach to the pre-distortion and equalizer circuitry used in its backplane transceivers (see the review in this week's i/oZONE.
How It Works
Quellan spent a good deal of time deciding on exactly where to partition
the digital and analog functions of its canceller, and the best ways to
implement these functions in vanilla CMOS. The result is a rather clever
collection of multi-pole FIR
filters, phase adjusters, and other components that can be steered by the
digital control logic. For a bit-head (former digital engineer) like myself,
some of the analog intricacies of the Nx600 were a bit of a challenge to
get my head around, but thanks to the patience of the Quellan technical
staff, I think I can provide you with a reasonable first-order explanation
of its operation.
The chip physically samples both the channel whose signal it's cleaning up (known as the "victim") and the adjacent channel creating the crosstalk (the "aggressor") using high-impedance taps. The chip's delay (phase control) circuitry creates an accurate time window within which to sample the aggressor signal so you can look at whatever 6-inch section of the signal path you want.
Configuring the Nx600 to cancel cross-coupled noise takes several steps. First, it develops a model that describes the coupling between aggressor and victim channels by stimulating the aggressor channel (during a training period at power on) with a burst of a known signal from an on-chip data generator. The sample is passed through a pole zero (PZ) filter to model how it is attenuated as it travels across backplane, and a FIR to shape the signal to approximate the frequency response of the channel. In operation, the aggressor channel is fed through these filters to create a cancellation signal.
Once the cancellation signal is generated, it's combined with the victim signal. A digitally-controlled fine-tuning delay circuit sits ahead of the combiner to precisely align the cancellation signal for minimum interference. There are several other ways for applying the circuitry to a particular problem, most of which can be managed by a simple FPGA-based state machine, and perhaps the host system via its serial I/O port.
While the Nx600 contains a single canceller, the chips can be run in parallel to cancel noise from two adjacent aggressor channels. You can also put two cancellers on the same channel to deal with two different signals originating from different portions of the line. This would typically be the NEXT from connector reflections at the near end as well as the CoXT from the PCB tracers.
Hardware Time
With a semi-coherent grasp of how the Nx600 works, it was time to take a
look at the hardware itself while it was being demonstrated at a conference
in Palo Alto. Quellan's test rig
was relatively straightforward, if not unwieldy. The tall stacks of signal
generators, oscilloscopes, and bit error rate testers (BERTs) filled one
end of the hotel suite and dwarfed the test backplane and the RT500 demonstrator pod that simulated the functionality
of the two Nx600 chips.
The RT500 pod used for the demonstration is built up using Quellan-built prototype discrete equalizer chips and a collection of other commercial components that mimic the functionality of a pair of canceller channels. Quellan will be making it available to potential customers who want to experiment with the canceller technology a few months before they can get a hold of the chips themselves.
Quellan's demonstration used a standard XAUI-style backplane with a signal generator running a 32K PRBS signal over 20 inches of FR-4 backplane. The receive end was connected up to a fast digital oscilloscope and a BERT. There were also two other pattern generators, each driving an adjacent set of traces to simulate a pair of aggressor channels. We first ran the signal alone at 6.125 Gbit/s and got a BER of somewhere above 10E-12. When the aggressor signal was introduced, the eye pattern immediately began to close up and the BER fell to around 10E-6 or worse.
Once the canceller was activated, the eye pattern recovered and produced a steady link with a BER close to, or equal to the unimpaired signal. We also sat and played a bit with adjusting tap parameters and the phase of the window location, but the original values seemed to produce the best results.
Conclusions
From what I can see, if the Nx600 matches the performance of its prototype,
it will be a very useful tool in mining hidden bandwidth from the backplanes
of the large installed base of networking equipment. There are some questions
in my mind about how closely the integrated chip can match the discrete
prototype's performance, but I think that any problems with substrate coupling
or other noise injection can be overcome. If anything, the chip may eventually
have a somewhat better frequency range due to the reduced parasitic loads
in the connections between functional elements.
My only other concern is with respect to the logistics of using the Nx600 in applications where line cards will be plugged into random locations on a backplane. Since the canceller is intended to only be used in the handful of "problem channels" sprinkled pretty much at random across a backplane, a line card would either have to be designed to go into a specific slot with cancellers attached only to the channels that need them. If a line card were designed to be poked into a slot at random, it might have to carry a bunch of extra cancellers to cover every set of pins likely to encounter a noisy signal at any slot position. At $20/chip this could easily add $60 or more to the BOM of a line card.
Conversations with my partner Paul led us to speculate that the problem could be at least partly solved with a high-frequency crosspoint switch (often used in broadcast equipment) and some simple logic that could take a minimum number of cancellers and connect them up to the appropriate pins according to which slot the card found itself in. I'm not 100% sure if this is a practical solution or if you'd just have to pay a penalty to allow your cards to be slot-independent.
From what I've seen, Quellan has just added another tool to the arsenal designers will use to increase the capacity of existing backplanes. It looks like the Nx600 can be used to augment the cost-effective SerDes products that are arriving on the market by providing a cost-effective drop-in fix for only the worst-case traces in a backplane where pre-emphasis and equalization alone are not sufficient to support 5-6Gbits/s data rates.
The prototype's FIR ICs appear to run well with some margin at 6 Gbit/s in the same 0.18 micron CMOS process used to make the Nx600 chip. This means that moving to 0.13 micron should allow a move up to 10 Gbit/s, and 20 Gbit/s look very do-able with a move to 0.09 micron technology. Quellan says that its CMOS design adheres to conservative rules that should make integration as an IP core relatively easy if the market demands it. All this gives Quellan's cancellation technology the "legs" it will need to address the evolving needs of the backplane market, as well as the other areas they've hinted at.
|
It is a relief to come across at least one start-up this year who is not trying to sell technology rather than a solution. Quellan has correctly identified that the two forms of crosstalk on a backplane -- the inevitable from connectors never designed to operate at GHz frequencies, and the cross-coupling crosstalk created by traces going where they shouldn't go, but often have to -- are a major headache for a SerDes design. There is no point throwing a signal at an equalizer if there is no real signal available for extraction and equalization, and at the frequencies being talked of here there will be many occasions where there just has to be a correction applied before equalization is attempted.
Saying that this correction could take place using a digital solution is like the cries of the movie director, "We'll fix it in post." You might be able to do something to mask the effects but you're not going to rid yourself of the problems. There are all sorts of cancellation techniques for audible noise: Noise canceling headphones, noise canceling of engine noise in turboprop airplanes like the Dash-8, noise canceling refrigerators in Japan. The same techniques are adaptable to any frequency providing you engineer it right. And those are analog techniques because whatever the nature of the electrical signals passing through a via, the signals are analog. What the signals are doing, whether they are modulated or not, is immaterial: They are analog signals. Once that small but dreadfully important fact gets through to the digital engineer all pretense at applying a digital solution should stop.
We have been using crosstalk and noise canceling techniques in electronics at frequencies higher than these backplane rates for over 15 years. They are straightforward, relatively easy to engineer and reliable for fixed operating conditions. Quellan is applying those techniques in CMOS and to an analog engineer, like myself, it is all entirely logical. No smoke and mirrors, no magic, just logical. There are limitations, of course, in particular choosing a frequency range that will be optimized by the techniques. Today what I know of the backplane market means looking for optimization up to about 5 GHz. The only other little gritty thing here is the data generator technique being used to judge the effects from the aggressor to the victim: A static clean signal like that will never perfectly match the actual effects that the real signal will create, but I cannot see in my mind any other way of doing it to better than this first order; the results tend to indicate that there is a big enough improvement that for all practical purposes it is not necessary to go to any further complication.
Quellan has already been able to persuade a few systems manufacturers of their philosophy but there will no doubt be a number of stubborn hold-outs who believe they can solve things for themselves -- to whom I wish the best of luck. And the SerDes manufacturers should embrace Quellan as helping them in maximizing their product potential by relieving them of a heavy analog load most seem disinclined to recognize or accept.
From a business perspective Quellan realizes that this market is not going to be around for ever; when the industry improves sufficiently (4 years, maybe?) for people to throw away their old racks and start afresh there may be OEM opportunities for small segments, but not much. The company has, however, identified other arenas that will readily accept their technology on an ongoing basis so that they will be able to expand out, probably into some arenas with considerably more healthy ASPs. One of the other things that the company has done right is to get a semiconductor vendor -- with fabs -- to invest in them. National Semiconductor is that investor and while the investment is only in single digits it means that Quellan will have a lot more clout on TSMC, who is making its silicon, than many other fabless companies out there. And capacity in the fabs will get a lot tighter than it is today.
I would measure Quellan by saying that it has done its initial crazy thinking, has realized a problem area, and then has gone for the solution of that problem with a straightforward, readily engineered bland CMOS solution. Not at all like another half-dozen start-ups I have seen this year who are still wondering who wants their technology.