Allan Herriman <allanherriman@hotmail.com> wrote:

(snip, I wrote)
>> 100baseTX uses MLT3, which reduces the bandwidth, but requires
>> a better S/N. There are three different voltage levels on the wire,
>> instead of two.
 
>> I don't know of a good reference for the PAM5 coding for 1000baseT.
 
> 802.3?  ;)

Yes, but I meant ones understanable by ordinary engineers.
 
> It's trellis coded.
> Whilst it's tempting to say that the PAM5 coding has five levels, each 
> four dimensional symbol actually consists of five levels on each of the 
> four pairs.  There are 5^4 = 625 possible symbols used to transmit 8 bits 
> of user data at 125MBaud.  The 625 symbols are divided into two sets that 
> are used every other symbol to improve the hamming distance.
 
> I am not aware that there is any spectral control at the coding layer, 
> i.e. there isn't a null at 0Hz (as would be expected with something like 
> 8B10B).

Seems to me that 8B10B goes asymptotially to zero, but close enough
not to worry about it.  More important is the upper edge of the band,
which both 100baseTX and 1000baseT keep below 100MHz. At 125Mbaud,
there should be signal to 62.5MHz, and decreasing above that.

From the DSP side, using more levels allows for a reduced bandwidth,
though requires more S/N. As I understand it, the coding, especially
the increased Hamming distance noted above, recovers the lost S/N.

But also interesting for DSP is the echo cancelation required, due
to using all pairs both directions. When you consider the attenuation
over 100m of cable, the receiver has to extract the smaller signal
coming in, while a strong signal is going out. Reflections from
impedance discontinuities, such as connectors, also have to be
subtracted.  At 125Mbaud, the pulses are around 2m long on the cable.  

As far as I know, it doesn't work at all without DSP.

-- glen

On Fri, 17 Apr 2015 17:08:07 +0000, glen herrmannsfeldt wrote:

> Tim Wescott <seemywebsite@myfooter.really> wrote:
> 
> (snip regarding 8b10b coding)
> (then I wrote)
>>> The fiber ethernets use it, 1000baseSX and such.
>  
>>> They use codes with either 5 ones and 5 zeros, or pairs of codes with 
4
>>> ones and six zeros, or the other way around. That, and a running
>>> disparity, allows one to keep close enough to balanced. They also 
don't
>>> use codes that will result in more than four consecutive ones or 
zeros. 
>>> After the code pairs for 256 data values are taken, there are still 
some
>>> left for other uses, such as an idle code when no data is being sent.
>  
>>> There are 252 codes with five 0s and five 1s, 210 pairs of codes with
>>> four and six, enough to remove those with too many zeros or ones in a
>>> row.
>  
>> That sounds an awful lot like regular 10base-T and 100base-T.
> 
> 10baseT is Manchester on UTP. Different collision detect from
> coax, but the signals are the same. 
>  
>> I assume that the multi-level coding schemes (why does my brain 
>> want to replace "coding" with "marketing"?) also have some sort 
>> of zero DC  property built into them.  I just don't know.
> 
> 100baseTX uses MLT3, which reduces the bandwidth, but requires
> a better S/N. There are three different voltage levels on the wire,
> instead of two.
> 
> I don't know of a good reference for the PAM5 coding for 1000baseT.

802.3?  ;)

It's trellis coded.
Whilst it's tempting to say that the PAM5 coding has five levels, each 
four dimensional symbol actually consists of five levels on each of the 
four pairs.  There are 5^4 = 625 possible symbols used to transmit 8 bits 
of user data at 125MBaud.  The 625 symbols are divided into two sets that 
are used every other symbol to improve the hamming distance.

I am not aware that there is any spectral control at the coding layer, 
i.e. there isn't a null at 0Hz (as would be expected with something like 
8B10B).
I may have misinterpreted the specification though.

Regards,
Allan

Tim Wescott <seemywebsite@myfooter.really> wrote:

(snip regarding 8b10b coding)
(then I wrote)
>> The fiber ethernets use it, 1000baseSX and such.
 
>> They use codes with either 5 ones and 5 zeros, or pairs of codes with 4
>> ones and six zeros, or the other way around. That, and a running
>> disparity, allows one to keep close enough to balanced. They also don't
>> use codes that will result in more than four consecutive ones or zeros. 
>> After the code pairs for 256 data values are taken, there are still some
>> left for other uses, such as an idle code when no data is being sent.
 
>> There are 252 codes with five 0s and five 1s, 210 pairs of codes with
>> four and six, enough to remove those with too many zeros or ones in a
>> row.
 
> That sounds an awful lot like regular 10base-T and 100base-T.

10baseT is Manchester on UTP. Different collision detect from
coax, but the signals are the same. 
 
> I assume that the multi-level coding schemes (why does my brain 
> want to replace "coding" with "marketing"?) also have some sort 
> of zero DC  property built into them.  I just don't know.

100baseTX uses MLT3, which reduces the bandwidth, but requires
a better S/N. There are three different voltage levels on the wire,
instead of two.

I don't know of a good reference for the PAM5 coding for 1000baseT.

-- glen

On Fri, 17 Apr 2015 13:14:30 +0000, glen herrmannsfeldt wrote:

> Tim Wescott <seemywebsite@myfooter.really> wrote:
> 
> (snip, someone wrote)
>>> Ethernet uses 8b/10b encoding to remove DC offsets.
>  
>>> http://en.wikipedia.org/wiki/8b/10b_encoding
>  
>> Is it used in Gigabit?  Gigabit uses a multi-level coding scheme, so
>> just encoding the raw data as 8b/10b wouldn't necessarily result in a
>> zero-DC signal.
> 
> The fiber ethernets use it, 1000baseSX and such.
> 
> They use codes with either 5 ones and 5 zeros, or pairs of codes with 4
> ones and six zeros, or the other way around. That, and a running
> disparity, allows one to keep close enough to balanced. They also don't
> use codes that will result in more than four consecutive ones or zeros. 
> After the code pairs for 256 data values are taken, there are still some
> left for other uses, such as an idle code when no data is being sent.
> 
> There are 252 codes with five 0s and five 1s, 210 pairs of codes with
> four and six, enough to remove those with too many zeros or ones in a
> row.

That sounds an awful lot like regular 10base-T and 100base-T.

I assume that the multi-level coding schemes (why does my brain want to 
replace "coding" with "marketing"?) also have some sort of zero DC 
property built into them.  I just don't know.

-- 

Tim Wescott
Wescott Design Services
http://www.wescottdesign.com

Tim Wescott <seemywebsite@myfooter.really> wrote:

(snip, someone wrote)
>> Ethernet uses 8b/10b encoding to remove DC offsets.
 
>> http://en.wikipedia.org/wiki/8b/10b_encoding
 
> Is it used in Gigabit?  Gigabit uses a multi-level coding scheme, so just 
> encoding the raw data as 8b/10b wouldn't necessarily result in a zero-DC 
> signal.

The fiber ethernets use it, 1000baseSX and such. 

They use codes with either 5 ones and 5 zeros, or pairs of codes
with 4 ones and six zeros, or the other way around. That, and a
running disparity, allows one to keep close enough to balanced.
They also don't use codes that will result in more than four
consecutive ones or zeros.  After the code pairs for 256 data values
are taken, there are still some left for other uses, such as an
idle code when no data is being sent.

There are 252 codes with five 0s and five 1s, 210 pairs of codes
with four and six, enough to remove those with too many zeros or
ones in a row.


-- glen

Rob Doyle <radioengr@gmail.com> wrote:
> On 4/7/2015 3:35 PM, Rob Doyle wrote:

(snip on gigabit ethernet and transformer coupling)

>> Ethernet uses 8b/10b encoding to remove DC offsets.

>> http://en.wikipedia.org/wiki/8b/10b_encoding

(snip)
> Re-read the original question.
 
> One of the simplest (and earliest) ways to achieve a DC balance is to
> use Manchester encoding.
 
> http://en.wikipedia.org/wiki/Manchester_code
 
> Manchester encoding ensures that the average DC value is zero over any
> bit. 8b/10b ensures that the average DC value is zero over a sequence of
> bits. Therefore 8b/10b encoding has more low-frequency content and
> requires a transformer with more iron (bigger?) than a Manchester
> encoded signal. Manchester encoding has more high-frequency content (for
> fixed data rate).

Yes, but the transition rate is higher, so the bottom of the 
spectrum also moves up. It isn't how many zeros (or ones) are in
a row, but how long they take to send.

For 10baseT, they fit the transformers into a DIP package, smaller
than a 74LS00 used to be (and still is). 

Note that Manchester is an inefficient code. It requires a bandwidth
up to the data rate, which is twice as high as Nyquist figured out.

For those who forgot, the actual problem Nyquist solved was the 
inverse of the one commonly named after him. 

Keeping the lower frequency high enough for transformer coupling
is important, but, at higher data rates, keeping the high end down
is also important. With 8B10B, the transition rate is 1.25 times
the data rate, and the bandwidth half of that. At 250Mb/s/pair, 
that would still be higher than the 100MHz spec for Cat 5 cable.
The 8B10B is used for fiber ethernet, which doesn't have the same
problem with high frequencies, but does still have a DC problem.

1000baseT uses the PAM5 coding, which reduces the upper frequency 
even more, though requiring a higher S/N than simpler coding.
PAM5 uses five voltage levels, so allows log2(5) bits per symbol
(transition), and so a lower bandwidth.  log2(5) is about 0.3,
and even with the more complicated coding gets the bandwidth
below 100MHz.

> BTW I think 8b/10b was developed (and patented) by IBM for use in
> their disk drives where the magnetic flux transitions need to balanced.
> I think it's still part of the SATA disk standard.

> One of the many places where data communications and data storage
> technologies overlap.

-- glen

On Tue, 07 Apr 2015 15:35:03 -0700, Rob Doyle wrote:

> On 4/7/2015 2:55 PM, Usjes wrote:
>> Hi,
>>
>> I am interested in a baseband digital communications system (Gigabit
>> Ethernet). The signal spectrum is baseband and there are isolation
>> transformers at both ends of the channel (cat5 cable). So one of the
>> channel impairments that the transceiver must deal with is baseline
>> wander. ie if there are a long string of 0's (or 1's) the signal begins
>> to look like DC (0Hz) and this will not pass through the (highpass)
>> transformers. So I would like to know what the typical algorithms for
>> cancelling this baseline wander effect are and what the pros and cons
>> for each algorithm is. So I consulted a few textbooks on DSP and on
>> digital comms but I cant find any entry in the index for Baseline
>> Wander. There are references for all the other blocks you will
>> typically find in a baseband digital transceiver; equalization, echo
>> cancellation,
>> timing recovery, crosstalk cancellation, but nothing on BLW ! So I am
>> wondering is this because:
>> (i) Baseline Wander is too obscure a topic (it would not be relevant
>> for wireless systems for example) and so typical texts dont cover it ?
>> (ii) There is some other more common name for this effect that it will
>> be listed under in the index ?
>> If (i) is the case, does anyone know of any online reference that might
>> list some of the approaches typically used to compensate for baseline
>> wander in baseband digital comms ?
>>
>> Thanks,
>>
>> Usjes.
>>
>>
> Ethernet uses 8b/10b encoding to remove DC offsets.
> 
> http://en.wikipedia.org/wiki/8b/10b_encoding

Is it used in Gigabit?  Gigabit uses a multi-level coding scheme, so just 
encoding the raw data as 8b/10b wouldn't necessarily result in a zero-DC 
signal.

-- 

Tim Wescott
Wescott Design Services
http://www.wescottdesign.com

On 4/7/2015 3:35 PM, Rob Doyle wrote:
> On 4/7/2015 2:55 PM, Usjes wrote:
>> Hi,
>>
>> I am interested in a baseband digital communications system (Gigabit
>> Ethernet). The signal spectrum is baseband and there are isolation
>> transformers at both ends of the channel (cat5 cable). So one of the
>> channel impairments that the transceiver must deal with is baseline
>> wander. ie if there are a long string of 0's (or 1's) the signal begins
>> to look like DC (0Hz) and this will not pass through the (highpass)
>> transformers. So I would like to know what the typical algorithms for
>> cancelling this baseline wander effect are and what the pros and cons
>> for each algorithm is. So I consulted a few textbooks on DSP and on
>> digital comms but I cant find any entry in the index for Baseline
>> Wander. There are references for all the other blocks you will typically
>> find in a baseband digital transceiver; equalization, echo cancellation,
>> timing recovery, crosstalk cancellation, but nothing on BLW ! So I am
>> wondering is this because:
>> (i) Baseline Wander is too obscure a topic (it would not be relevant for
>> wireless systems for example) and so typical texts dont cover it ?
>> (ii) There is some other more common name for this effect that it will
>> be listed under in the index ?
>> If (i) is the case, does anyone know of any online reference that might
>> list some of the approaches typically used to compensate for baseline
>> wander in baseband digital comms ?
>>
>> Thanks,
>>
>> Usjes.
>>
>
> Ethernet uses 8b/10b encoding to remove DC offsets.
>
> http://en.wikipedia.org/wiki/8b/10b_encoding
>
> Rob.

Re-read the original question.

One of the simplest (and earliest) ways to achieve a DC balance is to
use Manchester encoding.

http://en.wikipedia.org/wiki/Manchester_code

Manchester encoding ensures that the average DC value is zero over any
bit. 8b/10b ensures that the average DC value is zero over a sequence of
bits. Therefore 8b/10b encoding has more low-frequency content and
requires a transformer with more iron (bigger?) than a Manchester
encoded signal. Manchester encoding has more high-frequency content (for
fixed data rate).

BTW I think 8b/10b was developed (and patented) by IBM for use in
their disk drives where the magnetic flux transitions need to balanced.
I think it's still part of the SATA disk standard.

One of the many places where data communications and data storage
technologies overlap.

Rob.

On Tue, 07 Apr 2015 16:55:58 -0500, "Usjes" <59494@DSPRelated> wrote:

>Hi, 
>
>I am interested in a baseband digital communications system (Gigabit
>Ethernet). The signal spectrum is baseband and there are isolation
>transformers at both ends of the channel (cat5 cable). So one of the
>channel impairments that the transceiver must deal with is baseline
>wander. ie if there are a long string of 0's (or 1's) the signal begins
>to look like DC (0Hz) and this will not pass through the (highpass)
>transformers. So I would like to know what the typical algorithms for
>cancelling this baseline wander effect are and what the pros and cons
>for each algorithm is. So I consulted a few textbooks on DSP and on
>digital comms but I cant find any entry in the index for Baseline
>Wander. There are references for all the other blocks you will typically
>find in a baseband digital transceiver; equalization, echo cancellation,
>timing recovery, crosstalk cancellation, but nothing on BLW ! So I am
>wondering is this because:
>(i) Baseline Wander is too obscure a topic (it would not be relevant for
>wireless systems for example) and so typical texts dont cover it ?
>(ii) There is some other more common name for this effect that it will
>be listed under in the index ?
>If (i) is the case, does anyone know of any online reference that might
>list some of the approaches typically used to compensate for baseline
>wander in baseband digital comms ?
>
>Thanks, 
>
>Usjes. 

There are a number of techniques that are used to mitigate DC-offset,
like scramblers, but ethernet and many optical systems use 8b/10b
encoding as Rob pointed out.


Eric Jacobsen
Anchor Hill Communications
http://www.anchorhill.com

---
This email has been checked for viruses by Avast antivirus software.
http://www.avast.com

Usjes <59494@dsprelated> wrote:
 
> I am interested in a baseband digital communications system (Gigabit
> Ethernet). The signal spectrum is baseband and there are isolation
> transformers at both ends of the channel (cat5 cable). So one of the
> channel impairments that the transceiver must deal with is baseline
> wander. 

Reminds me of a discussion once on comp.dcom.lans.ethernet.

UTP ethernet uses balanced signals to avoid common mode interference.
As long as it is balanced, it won't radiate and most signals that
would be picked up by a long cable also won't bother it.

The transformers are supposed to allow either 500V or 1000V
of common mode on the cable. It was suggested, and maybe even
tested, that a 120VAC common mode voltage could be added, and it
would still work.  Just be careful where you touch it.

-- glen