principle I agree that this would be a great test to add to the simulation
suite. However, it is inherently a time domain test, and simulating it would
take a looong time to get to a reasonable confidence level in a 1e-12 BER
measurement. Note that all the results that we are comparing are based on
frequency domain simulations (with the exception of the LDPC code coding gain
is why the Crane test (with the equalizers frozen) has endured as a good measure
of system tolerance to external noise.
not have a "Dove test" which is just like the Crane test except it uses
randomly spaced transient impulses of a finite amplitude?
in 1000BASE-T days, I provided a presentation where I measured noise on CAT-5
installations and one site in particular was located next to an 89.9MHz radio
transmission tower. I really expected to see a lot of 89.9MHz carrier on the
wire, but much to my surprise, found transient pulses due to what was likely
an air-conditioner that were much worse in amplitude. The amplitude of these
pulses varied from a few mV (which is approximately the Vpp of the 89.9MHz
carrier on the wire) to as much as 19mVpk.
model for such a noise source could be readily defined as randomly distributed
between 50 and 70Hz and from 1mV to 20mV. One could shape the
distribution if they wanted to. Also, the pulse spectra could be defined. I
don't have the data, but I believe I can dig up that presentation and perform
some FFT to get it if it would be useful.
the subject of the Crane test, it is just a simple way of evaluating system
robustness against noise. There is no reason why it cannot be evaluated with
worst case impairments ON, as well as OFF. But I believe the
cancellers/equalizers do need to be frozen, because the intent is to capture
the effect on the system of a transient noise event, that is not long
enough to allow the system to adapt to it. I agree with you that the
coding gain validity is in question, but since we are (fortunately)
comparing very similar systems (both PAM, and both using LDPC), I'm assuming
that this will affect both systems equally.
As was discussed last week, the bottom line performance goal
is 1e-12 BER which is determined most directly by the SNR at the decision
device ("slicer SNR") by the familiar BER vs SNR curves the task force has
been using so far. I'd much
rather see results presented in terms of slicer SNR than the more obscure
"input-referred RMS noise power". The margin is then simply the dB
difference between the "required SNR" and the slicer SNR.
Perhaps others could voice their preference.
As was also discussed, the "required SNR" for LDPC
codes must be determined by simulation. Error floor and BER
slope change issues inherent to many LDPC codes cannot be predicted and
simulations must be performed to demonstrate that 1e-12 BER performance is
possible from any given code. Extrapolations from 1e-9 or 1e-10 (or
even 1e-11) are not always a reliable predictor of required SNR for 1e-12
BER. We have not yet seen results presented that establish
the required SNR for the PAM8 case with the proposed LDPC (2048,1723) code
as we have for the PAM12 case.
Lastly, we didn't have time to discuss this in detail last
week but there is some concern about the applicability of the so called
"Crane" noise immunity test for these PHYs. Another
bottom line performance goal is for the *PHY + connecting hardware* to
pass legally required noise immunity tests. The noise immunity test
consists of modulated sinusoidal fields applied to an actual operating PHY
in a system. This PHY will still have all other noise sources and
will have it's cancellers and equalizers in normal operating mode.
As I understand it, the "Crane test" puts the PHY in the unrealistic
condition of 1) no other external noise sources and 2) equalizers frozen,
not adapting. The Crane test makes the further assumption that the
same coding gain predicted for white Gaussian noise will be valid for
sinusoidal noise - I don't believe I've seen presentations or literature
which backs this assumption up. I don't think we have had enough
discussion on the Crane test's advantages/disadvantages, options, and
relationship to reality to simply adopt it and use the results to base our
PHY architecture decision on.
Dr. Scott Powell
Senior Manager, Ethernet PHYs
default cancellation parameters and necessary parameters to create
transmit PSDs for both PAM8 and 12.
stds-802-3-10gbt@IEEE.ORG [mailto:stds-802-3-10gbt@IEEE.ORG] On Behalf Of Sailesh
Sent: Friday, July
16, 2004 9:01 PM
Subject: [10GBT] Proposed PAM8 vs.
PAM12 resolution process
I would like to propose the
following process for resolving the robustness of PAM8 vs. PAM12 towards
1. Compute the Optimum DFE
SNR Margin for PAM8 and PAM12 using solarsep_varlen7a.m for Models 1 and
3 using default cancellation parameters and -150dBm/Hz background
2. Compute the
input-referred RMS noise power at the slicer by integrating the residual
noise in the Optimum DFE solution. I volunteer to add this code to
solarsep_varlen7a.m unless someone else wants to do
3. Compute the
input-referred external noise power that can be tolerated for a BER of
1E-12 for both systems using the results from (1) and (2) above. I
volunteer to add this code to solarsep_varlen7a.m unless someone else
wants to do so.