Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Feliz Haridad




Rich,

At least the circles are getting a bit tighter.  See
below for details.  It is really starting to snow here.
Time to break out the skis and snowshoes.

----------------------------------------------------------

>
>Pat,
>
>My outstanding MAS proposal is PAM5x4 @ 5 Gbaud (2.5 GHz) to transport 10
Gbps
>of Ethernet data.

------------------------------------------------------------

Most communication texts would recommend a bandwidth of .7xBR
in order to optimize SNR.  What is the reasoning behind your
choice of .5xBR?  Your proposal of 5Gbaud would result in a
bandwidth of 3.5GHz nominally.

-------------------------------------------------------------

>I don't suggest going to 10 levels (I assume this is what you
>mean by PAM10) due to the large SNR degradation while at the same time not
>pushing the technology enough (I assume that by 1 Gbit you mean 1 Gbaud).
>Besides, 10 bits gives you 3+ bits per baud, and assuming that you use the
+ for
------------------------------------------

What I meant to say was 1Gbit or 1.25Gbaud
as we have in GbE today.
------------------------------------------
>Forward Error Correction, the 3 bits left over times 1 Gbaud yields a data
>transfer rate of 3 Gbps. This falls far short of the required P802.3ae 10
Gbps
>data transfer rate requirement.
>
>MAS is laser wavelength independent. There are no problems with laser
safety at
>850 nm that I know that are MAS specific. I don't understand what your
concern
>is on this issue?
>

-----------------------------------------------------------

I am not advocating PAM10 or any other multi-level signalling 
scheme.  It is useful as an in the example I mentioned to 
illustrate the greater availability of power at 1300nm.  The
accessible power level at 1300nm is greater than at 850nm by
an order of magnitude.  This would help in overcoming the SNR
degradations occasioned by MAS.  

The bit rate calculations at PAM10 and 1.25Gbaud or your
preference, PAM5 and 2.5GHz, both require 4 fibers to achieve
something like a 10Gb/s single channel.  It really is a question 
of what data rate one is comfortable with.  The laser safety
issue I am addressing is for the 850nm case only.

I will go through the numbers.  At present, we have -17dBm
sensitivity for an 850nm receiver.  The bandwidth is specified
at .7xBR = 875MHz.  In your proposal, the bandwidth is 3.5GHz
in an apples-apples comparison.  This should give us something
like a 6dB electrical SNR degradation, or 3dB optical.

This would put us at -14dBm for the reciever sensitivity.  At 
present, our minimum transmit power is -10dBm.  In order to
maintain a 7dB link budget, we would need to raise this minimum
power to -7dBm.  The IEC safety limits we have adopted require
the maximum power radiated to be less than -4dBm.  Because all
of the light which exits from the transmitter port is not 
coupled, the practical limit for the maximum power coupled into
a fiber is something like -5dBm.  Most transceiver manufacturers 
typically guardband like this to ensure compliance.

Therefore, the range for coupled optical transmit power is only
a narrow range of 2dB.  This represents a drastic reduction in
transmit coupled power limits for most of the optical transceiver
manufacturers.  If we were to discuss 1300nm as an option, there
would not be any issue with raising the transmit power limit.
I hope I have made this issue clear enough.

----------------------------------------------------------------

>Hari IS a chip-to-chip interface. It makes no sense to use Hari between the
>SerDes and the MAC unless the SerDes you're talking about is part of the
>transceiver and the other side of the SerDes is in the MAC (integrated
MAC/PHY).
>Pictures always speak 1000 words. Please allow me to illustrate the potential
>elements (chips and such) in a 10 GbE LAN PHY:
>
>+-----------+ XGMII +--------------+
>|           +------->              |             +------------------+
>|           |   .   |           E S|    Hari     |S (E)             |
>|  10 GbE   |   .   |  10 GbE   n e+------------->e (n) Transceiver |
Medium
>|           |  36   |           D r+------------->r (D)   Module   
>+------------>
>|    MAC    |   .   |  PCS/PMA  e D+------------->D (e)   (PMD)     | 1-4
fibers
>|           |   .   |           c e+------------->e (c)             |
>|           |   .   |             s|  FR-4 PCB   |s                 |
>|           +------->              |   Traces    +------------------+
>+-----------+ short +--------------+    <=20"
>
>Figure 1 - Location of Hari, MAC and discrete PCS/PMA Chip within a 10 GbE
>Device
>
>+--------------+
>|              |               +------------------+
>|           E S|     Hari      |S (E)             |
>|  10 GbE   n e+--------------->e (n) Transceiver |   Medium
>|           D r+--------------->r (D)   Module    +------------>
>|  MAC/PHY  e D+--------------->D (e)   (PMD)     | 1-4 fibers
>|           c e+--------------->e (c)             |
>|             s|   FR-4 PCB    |s                 |
>|              |    Traces     +------------------+
>+--------------+     <=20"
>
>Figure 2 - Location of Hari and Integrated MAC/PHY Chip within a 10 GbE
Device
>

-------------------------------------------------------------------

The following picture illustrates what I have been suggesting
as an alternative which makes some sense.  It helps define the
question of where to partition a little better.

+--------+ XGMII +-------------+
|        +------->             |        +------+       +-- -----+
|        |   .   |          E S|  Hari  |S (E) |       | Trans- |
|10 GbE  |   .   | 10 GbE   n e+-------->e (n) |       | ceiver | Medium
|        |  36   |          D r+-------->r (D) |1 line | Module |
+--------|       |             |        |      |-------|        | 
|  MAC   |   .   | PCS/PMA  e D+-------->D (e) | 10Gb  |  (PMD) | 1 fiber
|        |   .   |          c e+-------->e (c) |       |        |
|        |   .   |            s| FR-4   |s     |       |        |
|        +------->             | Trace  +------+       +--------+
+--------+ short +-------------+ <=20"

Figure 1 - Recommended partitioning of a transceiver module.

+--------------+
|              |               +-----+        +--------+
|           E S|     Hari      |S (E)|        | Trans- |
|  10 GbE   n e+--------------->e (n)| 1 line | ceiver |   Medium
|           D r+--------------->r (D)|--------| Module +------------>
|  MAC/PHY  e D+--------------->D (e)| 10Gb   |(PMD)   | 1 fiber
|           c e+--------------->e (c)|        |        |
|             s|   FR-4 PCB    |s    |        |        |
|              |    Traces     +-----+        +--------+
+--------------+     <=20"

Figure 2 - Preferred partitioning of Hari and Integrated MAC/PHY Chip
            within a 10 GbE Device 

------------------------------------------------------------------------

>Note that one of the primary purposes of Hari are to reset the link jitter
>budget within the Transceiver Module (PMD) allowing the PMD-medium-PMD jitter
>budget to be independent to that of all components from each PMD back to it's
>respective PHY.

------------------------------------------------------

Why do you feel independence is such an important goal?
I believe the lowest total jitter should be the goal.

--------------------------------------------------------
>
>Given your answer, what interface do you then propose between the SerDes
back to
>the protocol chip (e.g. MAC, PCS, etc.)? The general requirements for this
>interface are: 
>
>- Interface reliability;
>- 10 Gbps data transport;
>- Protocol and PMD independence;
>- Low pin count to enable high port density protocol ASICs, removable
>Transceivers, reasonable routing of high-speed traces through FR-4;
>- Jitter budget independence from the link medium jitter budget;
>- CMOS friendly, at least to 0.25 micron;
>- FR-4 friendly;
>- Reasonable distances (to 20");
>- Scalability to higher speeds.
>

---------------
See above.
---------------

>- Interface reliability due to 8B/10B error control;
>- 10 Gbps data transport;
>- Protocol and PMD independence;
>- Low pin count to enable high port density protocol ASICs, removable
>Transceivers, reasonable routing of high-speed traces through FR-4;
>- Jitter budget independence from the link medium jitter budget;
>- CMOS friendly, at least to 0.25 micron;
>- FR-4 friendly;
>- Reasonable distances (to 20");
>- Scalability to higher speeds.
>
>As an architect, I always try to ensure that the whole system works. I
believe
>that you have only addressed a small and portion of my question
adequately. That
>is, you have addressed the 10 Gbps SerDes to PMD issue by backing the
"SerDes up
>directly against the optical transceiver". I believe that this is already the
>way it is. Now please consider the path all the way back to the protocol
ASIC.
>If your answer is that the latter interface is Hari, then I believe that we
>agree, since this is what I show in figures 1 and 2 above, and all your logic
>plus Hari logic is part of the PMD.
>

----------------
Again, see above.
-----------------

>
>Yes. It is clearly a question of partitioning. However, the partitioning
being
>proposed is not indiscriminate and is intended to provide a workable and
>interoperable architecture. The principal purpose of Hari interfacing the
PMD to
>the MAC/PHY is to separate the PMD-to-PMD jitter budget from the rest of the
>system. I consider this Hari attribute to be of paramount importance to the
>development of a successful and cost-effective 10 GbE standard.
> 

------------------------------------------------------------------

I would like you to explain why you feel the separation of jitter
budgets is so important.  Maybe you can point me to a section
in one of your presentations which makes this point better. 
It seems to me the lowest total jitter should be the overriding 
goal.

-------------------------------------------------------------------
>
>Your thoughts are not borne out by the datacom and communications industry in
>general. For example, multi-level signaling AND multiple channels have
been used
>is scaling Ethernet from 10 Mbits to 100 and then 1000. As a matter of fact,
>1000BASE-T uses PAM5, 4 pairs of wire and supports simultaneous full-duplex
>traffic over each wire pair. The same is true in most wireless and wired
>systems. Kestrel Solutions is offering an 0C-192 optical solution employing
>Frequency Shift Keying.
> 

--------------------------------------------------------------------

I will argue this point with you.  I believe there are more examples 
of true serial networks.  Comparing the numbers between the installed 
base of optical GbE over fiber and the number of copper multi-amplitude,
multi-wire GbE connections is instructive.  The preponderance of GbE
connections are true serial and are not skew compensated multi-wire 
systems.
 
----------------------------------------------------------------------
>
>Absolutely. Since there are many, many chip-to-chip connections in all
networks. 
>
>I don't really understand your question. Hari is a chip-to-chip connection
which
>enables the optimization of network link architectures. I don't view Hari
as a
>network.  
>
>
>CMOS gates are cheap, especially in light of the benefits provided by
Hari. Hari
>is a great example of the power of CMOS in addressing high speed network
>equipment design requirements. 
>
>

---------------------------------------------------------------

I believe in the HARI word striping versus byte striping debate
you prefer the byte striping because it uses fewer gates and less
power.  I thought the same logic might apply here as well. 

----------------------------------------------------------------
>
>

Feliz Haridad,

Pat Gilliland
patgil@xxxxxxxxxxx