Andrew, our DesignCon host for this panel discussion, asked me to spend ten minutes speaking to this audience about the future. I suppose I could have prepared a talk about the far, distant future. About the use of optical backplanes, or perhaps about some of the work going on with quantum-level devices. I might have even invented a kind of quantum-tunneling backplane structure for this talk.
But, thinking about what matters most to you, that is what structures are you likely to see in the future later this year, I decided to address something fairly practical, close-in, and which I believe has a better-than even chance of affecting your designs in the near term.
Just for background, let's take a minute to think about how the world of electronic communications has progressed.
If any of you read that wonderful book about Edison* (Israel, Paul, Edison: A Life of Invention, John Wiley & Sons, 1998) that came out recently, you know that the field of telegraphy was not static. As development progressed, engineers figured out how to work the lines as higher and higher speeds, eventually surpassing the speed of human comprehension several-fold through the use of paper-tape based high-speed recording and re-transmission hardware, and fancy multilevel transmission schemes involving both the polarity and amplitude of the transmitted signal.
That's typical for communications systems. At first, we use only a basic, dumb transmission scheme that realizes only a fraction of the raw communicating ability (the Shannon information-theoretic bandwidth) of the transmission channel. Then later, over time, the transmission and reception circuitry is gradually made more sophisticated to mine out more and more of the usable bandwidth.
Telephones followed a similar curve. When first conceived they had to go a few miles, and they had to use cheap, available wiring, so that established the analog bandwidth of the first systems at about 3 KHz. The first digital modems to operate over phone lines used a very simple FSK coding with a bit rate of about 300 baud. After a few years, engineers found ways to press the operating speed up towards the limit of analog bandwidth, and how to use multi-level signaling (up to 256 levels) to make optimal use of the great SNR available on short telephone connections, and the bit rate skyrocketed to its present level of about 56Kbps.
It's been the same way with LAN connections. Operating at shorter distances, the natural bandwidth of the cabling is much greater than miles-long telephone wires, so the first LAN systems came out at about 1 Mbps. Now we're up to 1 Gb/s. The latest Gb/s systems use 5-level PAM, and four wire pairs in parallel, to squeeze everything possible out of a 100-m category-5 link.
OK, now let's consider the world of PCB connections. What's happened here is that we are still in the "dumb link" mode of usage for these connections.
The natural bandwidth of a short PCB trace being several Gigahertz, we haven't needed anything more, at least not yet. But thinking about the future, and thinking about LONG connections (like backplanes) it's obvious what will happen next: more sophisticated use of the backplane traces, more sophisticated adaptive equalization, and multi-level signaling.
Figure 2 illustrates several important points relevant to signaling over a long PCB trace. The assumption here is that we are talking about a BIG backplane, something for a BIG piece of network hardware.
I'll make the leap of assuming we are talking about a multitude of serial point-to-point connections on this backplane, a fully-connected non-blocking configuration of some sort. The traces (including the backplane plus stubs on each card) are about 48 inches long (that's pretty long). The traces are 7-mils wide, implemented as FR-4 striplines. Now, we could debate all day long about the exact assumptions used, and the exact amount of loss, but that's not what I want to get across in this figure. What I want to get across is that at high speeds the loss is dominated by dielectric losses in the FR-4 material. In this region the loss (in dB) increases proportionally to the operating speed at a slope of 20dB per decade.
At the same time the noise from connectors, packages, and so forth (mostly crosstalk) increases at a rate of +20 dB/decade. The difference between these two effects, the SNR, is deteriorating at -40 dB/decade as we go up above a gigahertz. This is a very steep drop-off.
Those of you familiar with communications theory will recognize that a circuit afflicted with merely 20 or even 30 dB/decade of SNR deterioration is optimally filled with a binary signal, operating at the highest cycle frequency possible. That's what we've always done to date, and in a world dominated by skin-effect losses, where the SNR deterioration isn't too fast, binary signaling is the way to go.
On the other hand, when you are faced with a brick-wall communications channel, one with a finite bandwidth above which the SNR deteriorates miserably (like an analog telephone channel), the best this to do, to the extent that you have any excess SNR to work with, is to use multi-level signaling.
In the PCB world, the -40 dB/decade SNR slope produced by dielectric loss forms, in effect, a brick wall. The best way to increase bandwidth on long, dielectric-loss-afflicted channels is to use multilevel signaling.
In it's most simplistic form, the multilevel argument follows the form shown in Figure 3. Start with a binary system, then chop the signaling rate by a factor of B and try using 2B levels. You gain SNR at a rate of log(B)*40 (because of the 40 dB/decade SNR slope), but lose SNR proportional at a rate of only 20*log(2B-1) due to the number of levels.
Let's work an example.
Figure 4 illustrates a binary system operating at 2.5 Gb/s. (This is still over a PCB stripline 48 inches long, 7 mils wide, in FR-4.
Let's reduce the operating frequency by a factor of two, and plan to use four-level coding (two bits per baud).
With two bits, B=2, so we gain a factor of log(2)*40=12 dB in SNR due to the lower operating frequency. Let's use four levels. That makes N=4, with three spaces in between levels, so we have cut our slicing threshold spacing by a factor of three, which penalizes us by 20*log(3)=9.5 dB.
We gained 12 and lost only 9.5. Looks like the multilevel system will have an SNR advantage of about 2.5 dB.
Of course, there's a lot more to the analysis than that, as we need to include other improvements due to adaptive equalization, error-coding, and so forth. The net result I'd like you to remember is this:
When facing a brick-wall deterioration in SNR (like that produced by dielectric loss effects in PCB's), consider multilevel signaling.
What's required to actually do MAS?
- Multi-level transmitter with
- Multiple receiver thresholds and a
- Fancy clock recovery circuit.
To make it work in the practical world you will also need...
- Adaptive thresholds, and
- While you're doing that, you might as well add...
- Adaptive equalization - to fix problems with channel pulse response, and
- Scrambled line coding - to fix problems with EMI.
I'm not suggesting that multilevel signaling (or multi-amplitude signaling, MAS) is easy. Far from it. This is the sort of thing you do when you need to go faster and don't have any other alternative. But, given the incredible resources made available to us by improvements in chip technology, I can't think of a better use for all those hundreds of millions of transistors?
And think about some of the benefits -- especially from adaptive equalization. It can cure problems with bad connectors, trace stubs, and other artifacts. Maybe we are headed towards a day when we can spend less time worrying about the impedance of the channel, and minor reflections, and more time worrying about important issues like backplane protocols and system throughput.
What about EMI? It's better. Scrambling the channel, a provision often incorporated to insure the rapid convergence of adaptive-equalization algorithms, will spread the emissions spectrum of the data lines, providing huge benefits in EMI compliance. And what about clocks? Don't need 'em. Every serial receiver comes with it's own built-it clock-recovery subsystem. The clock is coded in the data. There's aren't any separate clocks. Scrambled, serial links are an EMI dream-come-true.
MAS is Already in Common Use
- Telephone analog modems - to 56 Kb/s
- Telephone DSL modems - to 1.5 Mb/s
- Gigabit Ethernet - to 1000 Mb/s
The technology to do high-speed multiple-amplitude signaling exists today. The first PCB applications will be large backplanes with substantial dielectric loss.
If any of you doubt the practicality of MAS, let me remind you that there was a time when people thought the sixteen-point QAM constellation in the 2400-bps V.22bis telephone modem was exotic and difficult to implement.
We are way past that point now.
The only factor holding back the development of MAS transceivers for backplane applications today is market demand. When sufficient demand exists for multi-gigabit serial transceivers that can operate at long distances, I'm sure MAS chips will be developed to fill that need.