Hi Pat, I think all of your arguments are valid, but to get the full picture, a few more considerations have to be added. As it was noted earlier, there is little doubt that for a lossy interconnect, which pushes the limit in terms of overall attenuation, and has a smooth roll-off, reasonable reflection loss, low crosstalk and low SSN, there is no need to consider much beyond the bit-rate frequency. Even in this case, if you want to model the channel performance accurately, you have to assume SOMETHING about the channel performance above the bit-rate frequency, otherwise the time-domain resolution of the calculations wont allow you to get accurate results (of course, what 'accurate' means, may depend significantly on the application's needs). But let me stress again, since the question was about channel modeling in general, we cant ignore modeling crosstalk, structural resonances and SSN. And that is even more important when the main channel has high attenuation. When we are at the typical limit today (20dB), it means that any crosstalk appears to be ten times bigger on the weak received signal. So I would argue that the actual limit is the sensitivity bandwidth of the receiver and not the bandwidth of the main received signal. Crosstalk, SSN and resonances will boost harmonics even if it happens after a low-pass filter channel, but it will be more devastating if the aggressor happens to be a close-by transmit signal whose edges are still sharper. In the old days some logic families had a glitch-filter feature: they were insensitive to pulses below a given duration (usually 1 nsec). I am not aware of a similar spec for SerDes receivers, and we can definitely not count on the receiver's bandwidth limitations in case of silicons, which can operate over multiple bit rates: since they have to support the highest bit rate, at lower bit rates they will respond to harmonics as well, even if they dont when they operate at their max speed. And when we add to this the usual large tolerance window of silicon features, if we guarantee that also the slowest silicon will respond properly to the input signal, a fast silicon will definitely be sensitive to higher frequency noise as well. Regards, Istvan Novak Oracle Zabinski, Patrick wrote: > The debate over what "rule of thumb" to use with regard to modeling > bandwidth has been going on before I joined the SI-List. Some folks say > the bandwidth should be based on edge rate, some say it should be based > on data rate. With the more recent discussion, I stepped back and > reconsidered my recollection of the impetus for the two most prevalent > guidelines, and I felt compelled to toss a few ideas in the air and see > where they land. I invite discussion/debate. > > I'll first start with high-speed serial links, where "high speed" by my > definition is where the data rate is high enough that some form of > equalization is requied. In other words, if equalization is not used, > then the eye is effectively closed and the receiver has no clue what's > coming in. Yeah; it's a fuzzy/vague definition, but it's the best I got > for now. > > I contend (for sake of argument) that this scenario does not care about > edge rate. In effect, the attenuation at the upper frequencies is so > high that any resemblance of an edge is absorbed in the channel. > Considering an X Gbps serial link, the "Nyquist" frequency (I hate that > term, but it's convenient) is X/2 GHz. Looking at all the SerDes I've > used, they commonly compensate for up to 20 dB attenuation at Nyquist = > X/2. > > Going back to Fourier expansion, the first harmonic that contributes to > the edge rate is 3X/2 GHz. Looking at it differently, I have never seen > a SerDes that provides any equalization at 3X/2 (anyone else?). I > believe there are two reasons for not doing so: 1) CMOS simply doesn't > have the bandwidth to address 3X/2 frequency without a huge penalty in > power; and 2) it is not needed (as demonstrated by all the SerDes out > there working today). > > For this particular scenario where active channel equalization is > required, I make the case that Nyquist (X/2) bandwidth is sufficient. > Buy that? > > The next scenario is for the traditional low-loss situation like a DDR > interface where the channel losses are present but basically leave the > waveform in good shape. An interesting aspect of these interfaces is > that they run much slower in raw serial speeds, but their timing > constraints are often more challenging (mostly due to the large > setup/hold times of the interfaces). To alleviate the timing > challenges, we need to squeeze out every picosecond of horizontal eye > opening we can, which results in speeding up the edge rates. > > Here, the data rate is certainly a factor in the overall eye opening, > but I've often been surprised at how much time we spend squeezing out > the last 10 ps to 20 ps of timing margin. To get these last few > piecseconds, we skip over the data rate (which is often fixed for DDR > interfaces) and work on the edge. > > Again, we can look at the data rate to determine how many odd harmonics > we need to preserve the edge, but someone smart once told me that the > associated bandwidth of these harmonics can be derived form the edge > rate. Yeah; 0.35/tr is not an exact number, but it gives us a rough > baseline to start with (I believe this equation came from test equipment > manufacturers). I seem to recall (don't have a reference handy) that a > 3 dB bandwidth of 0.35/tr captures something like 90% of the energy of a > digital signal. Seems like a lot, but it is not enough if you're > squeezing out the last few picoseconds from a timing-constrained link. > Instead, we often need to go 3X higher (~ 1/tr) or 5X higher (1.5/tr) to > preserve the speedy edge we need to make timing. > > So, I contend that in this scenario with low-loss, timing-constrained > links, that the modeling bandwidth should be based on the edge rate. > Buy that? > > Regardless of the minimum bandwidth is needed, I do strongly suggest > that all modeling go beyond the minimum. As indicated by several people > over the years, there are several "gotchas" associated with ignoring > what happens above the minimum bandwidth, where "minimum" is based on > several assumptions/premises that do not always apply. > > As indicated by previous recent postings, my mind is not all that clear > the past few weeks (that's another story...), so don't be shy about > pointing out the flaws in my arguments. > > Pat > ------------------------------------------------------------------ > > > > ------------------------------------------------------------------ To unsubscribe from si-list: si-list-request@xxxxxxxxxxxxx with 'unsubscribe' in the Subject field or to administer your membership from a web page, go to: //www.freelists.org/webpage/si-list For help: si-list-request@xxxxxxxxxxxxx with 'help' in the Subject field List technical documents are available at: http://www.si-list.net List archives are viewable at: //www.freelists.org/archives/si-list Old (prior to June 6, 2001) list archives are viewable at: http://www.qsl.net/wb6tpu