400GBASE-SR16
800 Gb/s
# lanes
16
10
8
4
2
1
FIGURE 1: The data center multimode speed roadmap shows the current standards (in black). Blocks with question marks
indicate areas where there are multiple possible outcomes that should be considered for the future.
TRENDS IN DATA CENTER ARCHITECTURES
Leaf-spine architectures are popular because they reduce
the number of “hops” from server to server, which
minimizes latency. One of the consequences of this
approach is that it requires more physical connections.
The physical layer becomes more important, because
it needs to accommodate a simplex, duplex or parallel
networking scheme. This means that sometimes the
design, implementation, testing and validation can get
fairly complex.
Further complicating this scenario is that as the rate
of change in network speeds has accelerated, optical loss
budgets are shrinking, and supported channel lengths
are shortening. As a result, there is less budget to work
with in the physical layer, so the connectivity used needs
to be much better. The good news is there are now more
choices in connectivity and media types; the bad news is
there are MANY more choices.
STRATEGIES FOR MOVING TO 100 GB/S
The Data Center Roadmap (Figure 1) shows that
customers are currently running anywhere from 10 Gb/s
to 100 Gb/s. While 40 Gb/s and 100 Gb/s have been on
the market for a couple of years now, most data centers
do not have a need to utilize the full 40/100 Gb/s as a
single connection. With one MPO connector plugged
40 I ICT TODAY
into a QSFP transceiver, a switch could be provisioned to
send a full 40 Gb/s or 100 Gb/s between devices. It can
also break down that 40 Gb/s speed into four 10 Gb/s
lanes or a 100 Gb/s down into ten 10 Gb/s lanes while
only utilizing one port on the switch. This allows for
more efficient switch port utilization, but it also creates a
more complex physical layer architecture.
Pertaining to the roadmap in Figure 1, 100G BiDi is
proprietary right now, but when the 802.3cm standard is
completed in 2019, there will likely be a standards-based
version available. There will also likely be MSAs around
200G. Currently the technology is available for 50G/
over SWDM, and OEMs are waiting for the demand to
increase. Likely, this will be an interim step before 400G
interfaces are cost effective. 100G/work is underway
and is expected somewhere beyond 2022.
The trend in the industry is to simplify cabling as
soon as the technology permits. This is driven by the
overall applied link cost (cost per Gb). For example,
100GBASE-SR10 has been replaced by SR4 and may soon
be displaced by SWDM4 or BiDi technology.
To support higher speeds, different techniques like
increased lane rates and more efficient modulation
schemes are being used. For example, above 25Gb/s lane
rate, PAM-4 modulation will replace NRZ. This allows for
doubling the number of bits per symbol.
100GBASE-SR10
40GBASE-SR4
40GBASE-SWDM4
400G-SWDM4?
100G-SWDM4
800G-SWDM4?
400GB-SWDM4?
200G-SWDM4?
1 Tb/s?
400 Gb/s
200 Gb/s
100 Gb/s
Lane rate > 10 Gb/s
25 Gb/s 50 Gb/s 100 Gb/s
Encoding > NRZ PAM-4