The factors that will drive to put CDR in to the module are the
- We are seeing about -35 dB connector NEXT in the SFP+ at 5.5 GHz
- Large port count SerDes or ASIC will not be able to do better
than XFI for jitter (TJ=0.3 UI) and SFI SerDes
must deliver <0.28 UI of jitter to the module internal pin
across 6-8" of FR4 trace.
- Very difficult to route 10 TX/RX traces have low crosstalk and
good SI (4x would be easier and
you may be able to set the SerDes good enough to meet SFI
requirements not using CDR)
In case of 10X the cost will increase to 6x the cost of Xenpak due to
I also agree with Mike assessment that cost of 4x10G or 10x10G parallel
solution would be
respectively about 4x and 10x the cost of single 10G XFP. XFP cost is
about 1/2 the cost of
Xenpak and SFP+ is about 1/2 the cost of XFP.
If the 4x or 10x solution need CDR in the module then it will be very
difficult to come close
to cost of SFP+ per lane.
Paul Kolesar wrote:
thanks for the critique.
I think the confusion is a result of
two factors: the translation between the various form factors of
(e.g. SFP, SPF+, XFP, XENPAK), and time frame of the comparisons.
Kipp stated that presently the cost of 20G (4x5G) QSFP was about the
as XFP (or 10 x SFP cost), declining to 4 x SFP cost with volumes
time). I was basing my initial costs on the present comparison. It
would be useful for future analysis if the relative costs of these form
factors containing the same optical elements were made known. Scott
provided the touch stone for the 4G-SFP and 10G-XFP. Others that
would be useful include 10G-SFP+ and XENPAK.
My analysis of the cost of 4x
cabling with duplex (LC) connectors agrees with your assessment that it
is roughly the same as 1x 12-fiber cabling with array (MPO) connectors.
This can be seen in my cost comparison for the LAG scenario.
I realize that the cost of
I/O chips is a significant cost driver. I can accept that the factor
of 2 that I applied as the difference between the 20G QSFP and 40G QSFP
might be high. But it does remain to be seen if a 10G-lane-rate QSFP
can fulfil the needs of the yet-to-be-defined solution without CDR, one
type of such chips. I think Ali stated that he thought CDR would
be needed for parallel optics. If you are correct, then the difference
between 4x 10G LAG and 40G parallel would be greater, as the cost of
40G parallel would be lower.
I do not understand your reference
mux/demux chips, since it is presumed that the electrical lane rate
the optical lane rate.
If you agree that the cost factor of
2.5 is appropriate for 40G to 100G, then perhaps my guesstimate that
cost of 5 XFI is roughly the same as 4 XENPAK is a source of error.
what I infer from the previous point on the 20G to 40G comparison, then
perhaps the number of XFI should be higher. Please clarify as this
material is being used in presentations for next week.
1300 East Lookout Drive
Richardson, TX 75082
I disagree with some of the
made below, in particular the difference between 40G (4x10G as one
and 40G LAG (4x10G independent streams) for parallel implementation.
As Scott Kipp said the cost of a 40G QSFP (which does not include any
is likely to be somewhat more than the cost of 4 10G SFP parts.
The fiber interconnect for the SFP's are duplex LC whereas the QSFP has
an MTP (I'm not sure what the relative cost of 4 duplex LC jumpers
an MTP jumper is, but I doubt the difference is significant.). For
the link itself the fibers are likely to be similar. In short I
think that any significant cost differences are related to the cost of
the IC's that are required to mux/demux the signals for either 40G
or whatever is needed for LAG. There is no factor of two difference
between these options and what difference there is may be in the other
On the relative costs of the
versus 40G for parallel solutions I think you are fairly close although
you may have over-estimated the difference in cost between the
for the PMD's. Assuming we are talking 10x10G and 4x10G parallel
solutions I would expect that the costs would be similar per 10G
with the 4x10G implementation being somewhat higher on that metric.
are complexities in having Tx and Rx in the same package, and the total
number of components for the two solutions is smaller per 10G stream
the 100G). My estimate is that the relative costs of the two
PMD's would be between 2x and 2.5x larger for the 100G. (ie cost
per bit is either equal or somewhat less for the 100G solution which is
what you assumed).
Director Transceiver Engineering
1480 Arthur Avenue
Tel 303 530 3189 x7533.
From: Paul Kolesar
Sent: Tuesday, April 10, 2007 10:31 AM
Subject: Re: [HSSG] 40G MAC Rate Discussion
I'm returning to an item in the discussion from April 6, where Matt
> - I'd like to hear a comment / perspective from the fiber
> manufacturers on the utilization rate of the ribbon fiber strands.
> For a 4x10G MMF approach presumably 8 strands in the 12 ribbon
> be used 4 for TX and 4 for RX. For a 10x10G MMF approach it
> 2 @12 with 10 @ Tx and 10 @ RX. What does this do to the cost
> usage rate metrics of MMF cabling?
Some parts of my response to your inquiry are straight forward and easy
to provide without cause for concern. More in depth response on cost
gets tricky and I do not want to cross over the line our chair has
us of. So I have modified it a few times. As this material
developed, it appears that it could be the basis for a presentation.
it is acceptable and useful.
For parallel MMF connectivity the cost of the cabling for 4x10G will be
half that of 10x10G simply because 12-fiber cabling is used for the
and 24-fiber cabling is used for the latter. Also the patch panel
density (circuits per unit area) will be double for the former compared
to the latter. This may also be the case for active equipment panel
density. They both use the same cabling componentry, but the latter
requires twice as much.
The effect of this difference on the total cost of the system is
with the projected cost difference between the 40G and 100G PMDs, as
On April 9 Scott Kip wrote:
> I have heard that the cost of a 5 Gig QSFP are
> comparable to the cost of a single XFP or approximately 10 4-Gig
> If the QSFP rolls out in relatively high volumes, the QSFPs are
> to approach the cost of 4 SFPs at comparable speeds.
If this relationship is true and from what I know of such things, then
the cost to the end user of an installed, tested, warranted,
12-fiber, MM structured cabling link, consisting of a cable terminated
at patch panels plus array-terminated cords on each end, is in the
of today's cost of a pair of 20G QSFPs to an OEM (e.g. server, switch,
router maker). In the analysis that follows I'll assume that the
future 40G QSFPs are 2x the cost of 20G QSFPs (i.e. 2x XFP cost), and
100G PMDs are 5x XFP cost (trying to be consistent with
where 100G cost = 4x XENPAK cost). In order to compare cabling and
PMD costs at the end user level, a PMD cost adjustment factor to
for integration by the OEM must be applied. The same factor is applied
to all cases. The tabulations that follow should help make the basic
trends over time and between data rate choices clearer.
Using 2007 XFP cost to the OEM as the monetary unit, the basic end user
projected cost comparison is:
data rate 40G 100G
PMD cost 6 15
cable cost 1 2
total cost 7 17
Here cabling cost is ~1/7 (14%) to ~2/17 (12%) of total cost. 40G
total cost is 41% of the 100G total cost.
In about three to four years time (assuming both 40G and 100G MM PMDs
the same initial availability date) the cost is:
data rate 40G 100G
PMD cost 3 7.5
cable cost 1 2
total cost 4 9.5
Here cabling cost is ~1/4 (25%) to ~2/9.5 (21%) of total cost. 40G
total cost is 42% of the 100G total cost.
If the 40G PMD is available ahead of 100G PMD, then the 40G solution
be relatively lower in cost.
With the caveats of the above assumptions, the conclusion is that the
QSFP PMD should offer significant advantage in cost for the total
relative to the 100G PMD from both transceiver and cabling
The relative total cost ratio appears to stay fairly constant, even
while the cabling costs become relatively more significant. This
is because the ratio of the cabling costs is similar to that of the
PMD costs. To illustrate, if dominated completely by PMD cost
the ratio would be 40%, while if totally dominated by cabling cost
the ratio would be 50%.
If your metric of value is cost per gigabit, the two scenarios
equate. But if your metric is absolute cost, then one would not have
to pay for more than one needs with a 40G solution.
To offer another perspective, comparing to a LAGed 10G scenario:
data rate 10G 40G
PMD cost 3
cable cost 0.25 1
total cost/ckt 3.25 7
LAG factor 4 1
total cost 13 7
the 40G solution provides a very favorable cost comparison, at a cost
is a bit more than half of the 4xLAG 10G option.
As far as the other part of your question regarding usage rate metrics
for the cabling, as you can see from the above analysis, leaving four
fibers per cable (8 out of 12 active with 40G) instead of 2 (10 out of
12 active with 100G) does not have significant impact, since the user
the array-terminated cabling on a modulo 12 basis. Utilization
can make a difference if the modulo-12 cabling supports more than one
at a time. For example, it makes a cost difference if there are unused
fibers for applications where we deploy 2-fiber circuits over a
cable. The 12-fiber cabling gets amortized over one to six circuits
depending on utilization.
Both the 40G and 100G scenarios have degraded capacity utilization in
cable compared to 10GbE. Today a 12-fiber cable supports six 10GbE
circuits for a capacity of 60G. That same cable will support only
40G or 50G capacities for the 40G PMD or 100G PMD respectively.
But in both scenarios there are far fewer patch cords and port
to manage than for 10G links aggregated to match the comparative data
and the overall cost for either of the higher-rate solutions should be
very favorable compared their LAGed counterparts.
1300 East Lookout Drive
Richardson, TX 75082