100G and things to COM
FEC (forward error correction) changes the way we think about BER (bit error ratio/rate) in serial data links. If we can afford 7 errors in every set of 528 bits (i.e., for the Reed-Solomon coding used in 100 GbE, RS(528,514)), then we need a new way to characterize those network elements that cause errors.
In June, IEEE released the 100 GbE (gigabit Ethernet) specification, 802.11bj. 100GbE> accomplishes 100 Gbits/s by combining four 25.78125 Gbit/s lanes, the excess data rate accommodates overhead—we’ll call it 4x25 Gbits/s. Previous releases had neglected specification details for electrical signaling on cables and backplanes. That is, fiber optic transmission had been wholly spec'ed years ago but we just got specs for these blistering data rates on PCB (printed circuit board) and cables.
The backplane spec (called 100GBASE-KR4 in the spec) accommodates 4x25 Gbits/s over backplanes with total insertion loss up to 35 dB at 12.9 GHz, about half the data rate, and for cables (100GBASE-CR4), 4x25 Gbits's over shielded balanced copper cabling with reach of at least 5 m. Notice that backplanes are specified in terms of their loss and cables in terms of distance of propagation.
The spec has some new concepts that open up design options but can be confusing; for example, channel operating margin. COM (channel operating margin) provides flexible choices instead of straight-up pass/fail criteria or masks for s-parameters. The idea is to let engineers trade off between loss, reflections, and crosstalk. That last signal impairment, crosstalk, is at the root of most of this evil.
COM is the ratio of the signal amplitude to the noise amplitude given in decibels:
and the spec requires COM >3 dB, well, "require" is too strong a word. The COM spec is "
It looks innocent enough, right? Measure COM, require it be larger than a specified threshold and move along. But there’s a devil is in the denominator. The noise term has to include crosstalk, random noise and jitter (RJ), inter-symbol interference (ISI), and all the other acronymed culprits
COM test setup (Ransom Stephens).
Specific, though hardly restrictive models are used to extract COM. Three tap de-emphasis at the transmitter, called feed forward equalization in the literature, and both CTLE (continuous time linear equalization) and DFE (decision feedback equalization) at the receiver. To specify the intrinsic noise of a channel, it makes sense to use s-parameters or, equivalently, the impulse response of the channel, including its response to all three other crosstalk aggressors. Instead of worrying over the complication of constructing an actual impulse with infinite amplitude and infinitesimal width, a single bit-wide pulse is used; that is, instead of impulse response, COM uses the response of the channel to a single bit, the SBR (single bit response).
The noise term is extracted by using a channel’s SBR, including the aggressor SBR, combined with the best available transmitter and CTLE receiver equalization. COM is then the ratio of the signal voltage at eye center to the RMS (root mean square) noise after the CTLE.
Perhaps a more intuitive description is to derive the noise by combining the RMS noise of the transmitter, ISI, RJ, crosstalk, and amplitude noise something like this:
COM >3 dB is required; the design flexibility lies in how you divvy up the contributions from ISI, crosstalk, and transmitter noise.