Modem question

Eric Jacobsen <eric.jacobsen@ieee.org> wrote:
(snip)

>>> One probably could do that, but if he's looking for existing equipment
>>> his selections will probably be limited.   I don't know of any general
>>> purpose hardware that does that.   A few folks do it for satellite
>>> modems, where the spectrum is really expensive, but other than that it
>>> seems pretty rare.

>>I thought that all the high speed, 9600 bit/sec and higher,
>>modems did it.  

>>Gigabit ethernet, 1000baseT, does.
 
> Are they cancelling the back channel?   I.e., simultaneous Tx and Rx
> on the same channel with Tx cancellation at the Rx?    I think that's
> what Randy was asking about.    I wasn't aware that any wired
> standards did that, if they do.

Yes, simultaneous Tx and Rx.  

1000baseT uses four pair Cat5e cable, with all four going both
directions at 250Mb/s.  It is baseband, so no channel question.

I am not so sure by now when phone line modems started overlapping,
but definitely by 33.6kb/s, as half a 64kb/s channel wouldn't 
do that.

-- glen

Vladimir Vassilevsky <nospam@nowhere.com> wrote:
(snip)
 
> I doubt feasibility of high quality echo cancellation in 
> the bandwidth of tens of MHz.

More than that, I believe 100MHz.  1000baseT uses four channels
of 250Mb/s each.  I believe the coding keeps most of the energy
below 100MHz, but not much below.  Both directions simultaneously
on four pairs.

-- glen

>Vladimir Vassilevsky <nospam@nowhere.com> wrote:
>(snip)
> 
>> I doubt feasibility of high quality echo cancellation in 
>> the bandwidth of tens of MHz.
>
>More than that, I believe 100MHz.  1000baseT uses four channels
>of 250Mb/s each.  I believe the coding keeps most of the energy
>below 100MHz, but not much below.  Both directions simultaneously
>on four pairs.

I think its something like 125MHz bandwidth. Gig-E ports are pretty much
standard on PCs, but Gig-E switches still have a premium large enough that
most people are connected to a 100M switch port. Try connecting many
machines to a Gig-E switch and the ethernet chip cooks. You need to
substantially improve the cooling to achieve stable operation. All that DSP
takes real power. :-)

Steve

glen herrmannsfeldt wrote:

> Vladimir Vassilevsky <nospam@nowhere.com> wrote:
> (snip)
>  
> 
>>I doubt feasibility of high quality echo cancellation in 
>>the bandwidth of tens of MHz.
> 
> 
> More than that, I believe 100MHz.  1000baseT uses four channels
> of 250Mb/s each.  I believe the coding keeps most of the energy
> below 100MHz, but not much below.  Both directions simultaneously
> on four pairs.

That's relatively shallow echo cancellation, and over the media with 
tightly specified parameters. For example, echo canceling in DSL is more 
difficult.

VLV

steveu <steveu@n_o_s_p_a_m.coppice.org> wrote:
 
> All the modem standards up to V.32bis require that you tolerate a 8-10Hz
> shift between the sending frequency and the received one. This is to allow
> for FDM equipment in the channel, which was common until quite recently. I
> think V.34 leaves the issue loose, and V.90/V.92 are only for pure digital
> transmission networks.

(snip)

Well, V.90 and such are only "pure digital" in one direction, and
not really all that pure.  It has to go through one D/A converter
before it gets to you, and you have to be able to figure out from
the signal the sample points to decode the signal.  

In addition, there is possible echo from the reverse channel
mixed in with that.

-- glen

steveu <steveu@n_o_s_p_a_m.coppice.org> wrote:
(snip, I wrote)

>>More than that, I believe 100MHz.  1000baseT uses four channels
>>of 250Mb/s each.  I believe the coding keeps most of the energy
>>below 100MHz, but not much below.  Both directions simultaneously
>>on four pairs.
 
> I think its something like 125MHz bandwidth. 

http://www.cisco.com/en/US/tech/tk389/tk214/tech_digest09186a0080091a86.html

The symbol rate is 125M symbols/sec, but I believe coded to keep the
bandwidth down.  Cat5 and Cat5e are rated at 100MHz bandwidth.

> Gig-E ports are pretty much
> standard on PCs, but Gig-E switches still have a premium large enough that
> most people are connected to a 100M switch port. Try connecting many
> machines to a Gig-E switch and the ethernet chip cooks. You need to
> substantially improve the cooling to achieve stable operation. All that DSP
> takes real power. :-)

Five port gigabit switches are pretty affordable, I don't know
about cooling, though.  I believe it is one chip for the five
port switch.

-- glen

>steveu <steveu@n_o_s_p_a_m.coppice.org> wrote:
> 
>> All the modem standards up to V.32bis require that you tolerate a
8-10Hz
>> shift between the sending frequency and the received one. This is to
allow
>> for FDM equipment in the channel, which was common until quite recently.
I
>> think V.34 leaves the issue loose, and V.90/V.92 are only for pure
digital
>> transmission networks.
>
>(snip)
>
>Well, V.90 and such are only "pure digital" in one direction, and
>not really all that pure.  It has to go through one D/A converter

Have you ever seen a telephone path that is PCM all the way in one
direction, and has FDM links in the reverse direction? :-\ What I said is
accurate. V.90 is a modem that is only for use on pure digital transmission
networks.

>before it gets to you, and you have to be able to figure out from
>the signal the sample points to decode the signal.  

>In addition, there is possible echo from the reverse channel
>mixed in with that.

The echo cancellation for V.90 seems both harder and easier than the echo
cancellation for the slower modems.

Its easier in that you only get echo between the 2 hybrids at the analogue
modem end; there is no FDM frequency shift to worry about; there is no bulk
delay to figure out; the analogue modem end gets an echo which has no G.711
distortion. The modem at the digital echo must remove the echo of its 56k
signal from a signal distorted by the G.711 encoder in the outgoing
exchange. However, that is no worse than the problem it faces in a V.34
modem (since the upstream is, basically, half a V.34 modem).

It looks harder in that you really need to get the fine residue out of the
56k signal at the analogue modem end, to be able to get the full 56k rate.
I've never actually tried this, but from other echo cancelling work it
seems like that is going to take some effort, and demand a pretty linear
front end.

On Wed, 8 Sep 2010 03:25:01 +0000 (UTC), glen herrmannsfeldt
<gah@ugcs.caltech.edu> wrote:

>Eric Jacobsen <eric.jacobsen@ieee.org> wrote:
>(snip)
>
>>>> One probably could do that, but if he's looking for existing equipment
>>>> his selections will probably be limited.   I don't know of any general
>>>> purpose hardware that does that.   A few folks do it for satellite
>>>> modems, where the spectrum is really expensive, but other than that it
>>>> seems pretty rare.
>
>>>I thought that all the high speed, 9600 bit/sec and higher,
>>>modems did it.  
>
>>>Gigabit ethernet, 1000baseT, does.
> 
>> Are they cancelling the back channel?   I.e., simultaneous Tx and Rx
>> on the same channel with Tx cancellation at the Rx?    I think that's
>> what Randy was asking about.    I wasn't aware that any wired
>> standards did that, if they do.
>
>Yes, simultaneous Tx and Rx.  
>
>1000baseT uses four pair Cat5e cable, with all four going both
>directions at 250Mb/s.  It is baseband, so no channel question.

That part I was aware of, and the four pairs and their crosstalk are
essentially processed like a MIMO system.

I didn't know that 1000bT did simultaneous Tx/Rx.   The basic ethernet
MAC is TDD, and I didn't know how much that had changed.

>I am not so sure by now when phone line modems started overlapping,
>but definitely by 33.6kb/s, as half a 64kb/s channel wouldn't 
>do that.
>
>-- glen

Is the Tx cancellation done digitally?   I'm surprised if this is the
case since one would think it would make the ADC quite a bit more
expensive back in those days.


Eric Jacobsen
Minister of Algorithms
Abineau Communications
http://www.abineau.com

>On Wed, 8 Sep 2010 03:25:01 +0000 (UTC), glen herrmannsfeldt
><gah@ugcs.caltech.edu> wrote:
>
>>Eric Jacobsen <eric.jacobsen@ieee.org> wrote:
>>(snip)
>>
>>>>> One probably could do that, but if he's looking for existing
equipment
>>>>> his selections will probably be limited.   I don't know of any
general
>>>>> purpose hardware that does that.   A few folks do it for satellite
>>>>> modems, where the spectrum is really expensive, but other than that
it
>>>>> seems pretty rare.
>>
>>>>I thought that all the high speed, 9600 bit/sec and higher,
>>>>modems did it.  
>>
>>>>Gigabit ethernet, 1000baseT, does.
>> 
>>> Are they cancelling the back channel?   I.e., simultaneous Tx and Rx
>>> on the same channel with Tx cancellation at the Rx?    I think that's
>>> what Randy was asking about.    I wasn't aware that any wired
>>> standards did that, if they do.
>>
>>Yes, simultaneous Tx and Rx.  
>>
>>1000baseT uses four pair Cat5e cable, with all four going both
>>directions at 250Mb/s.  It is baseband, so no channel question.
>
>That part I was aware of, and the four pairs and their crosstalk are
>essentially processed like a MIMO system.
>
>I didn't know that 1000bT did simultaneous Tx/Rx.   The basic ethernet
>MAC is TDD, and I didn't know how much that had changed.

The MAC was never really TDD. the hub fabrics imposed the one at a time
restriction. When hubs gave was to ethernet switches almost everyone
started using their MAC in a full duplex manner.

>>I am not so sure by now when phone line modems started overlapping,
>>but definitely by 33.6kb/s, as half a 64kb/s channel wouldn't 
>>do that.
>>
>>-- glen
>
>Is the Tx cancellation done digitally?   I'm surprised if this is the
>case since one would think it would make the ADC quite a bit more
>expensive back in those days.

How else are you going to echo cancel in a practical design? The V.32 modem
at 9600bps was the first mass market PSTN modem to use EC, and basically
the same signal both ways (the initial training is asymmetric, and the two
directions using different scrambler polynomials, but they are essentially
the same). It was a darned expensive modem in the mid 80s, when it first
came to market. The ADC was not the tough part. The non-linear distortions
caused by G.711 coding in the channel, and the FDM issues I referred to in
other messages, where the real challenges. With V.34 the constellation has
such fine steps that the echo cancellation needs to be much tighter, and
the ADC does need to be better than the usual telephone grade. You still
have to cancel echoes through the G.711 channel, but a non-linear ADC gives
you problems suppressing what is normally the largest echo - the one on the
nearest channel segment.

Steve

On Wed, 08 Sep 2010 17:39:56 GMT, eric.jacobsen@ieee.org (Eric
Jacobsen) wrote:
>>>>Gigabit ethernet, 1000baseT, does.
>> 
>>> Are they cancelling the back channel?   I.e., simultaneous Tx and Rx
>>> on the same channel with Tx cancellation at the Rx?    I think that's
>>> what Randy was asking about.    I wasn't aware that any wired
>>> standards did that, if they do.
>>
>>Yes, simultaneous Tx and Rx.  
>>
>>1000baseT uses four pair Cat5e cable, with all four going both
>>directions at 250Mb/s.  It is baseband, so no channel question.
>
>That part I was aware of, and the four pairs and their crosstalk are
>essentially processed like a MIMO system.

Not really, at least not the way I understand MIMO. In 1000BT, only
near-end cross-talk (NEXT) is cancelled and FEXT is ignored. In 10GBT
over copper, FEXT cancellation is necessary and most implementation do
that and there are some published papesr on using off-diagonal channel
parameters (ie Hij where i!=j) as a means to estimate what data is
being transmitted for channel I.

>
>I didn't know that 1000bT did simultaneous Tx/Rx.   The basic ethernet
>MAC is TDD, and I didn't know how much that had changed.
>
Ethernet has been full-duplex since 100BTX days and even 10BT had full
duplex implementations. These mainly are FD over separate channels.
Starting with 100BT2 (which wasn't popular) and mainly with 1000BT,
simultaneous tx/rx over a single pair has been implemented. This is
done with a hybrid to mix the 4 wires to 2 and echo cancellation is
needed which is done by mainly digital means. 

>>I am not so sure by now when phone line modems started overlapping,
>>but definitely by 33.6kb/s, as half a 64kb/s channel wouldn't 
>>do that.
>>
>>-- glen
>
>Is the Tx cancellation done digitally?   I'm surprised if this is the
>case since one would think it would make the ADC quite a bit more
>expensive back in those days.

There are implementations which do detection in digital but partial
cancellation in analog ie before ADC to save ADC dynamic range. One
does bulk of the near echo cancellation in analog by an adder before
the ADC as the near echo can be 10-15 dB above received signal coming
from a far transmitter. Then there is a digital echo canceller to
reduce the residual echo below needed echo suppression range.
-- 
Muzaffer Kal

DSPIA INC.
ASIC/FPGA Design Services

http://www.dspia.com