Just like Buzz Lightyear and all the other toys from a certain very successful movie franchise, the data center industry never sleeps. Our insatiable appetite for bandwidth continues to challenge technology developers to keep pushing data faster and more efficiently than ever before. They are continually designing more engaging applications, processes, and products to enrich our social and professional engagements, de-humanizing the everyday mundane, complex tasks and enveloping them in the digital bubble we live in today. This is Demand and Supply 101. We demand it and boy, are they supplying it in such a way that it creates a super dependence we never knew we had. My wife often jokes she doesn’t know how I survived before the search engine!
What’s behind this drive for faster, more efficient network speeds? Basically it’s us. The human race. We get bored easily. We are constantly searching for new, more interactive ways to entertain ourselves, a more exciting and engaging method of communicating like never before and we expect it all to happen now. This instant. We don’t use the phone to talk to people anymore. Text messages are old school, and picture messages are rapidly becoming yesterday’s news. If you want to be down with the kids, it’s all about streaming ultra-rich content at every possible opportunity. Day and Night. But it goes further than this. We want to connect more devices to the information superhighway, we want improved analytics, we want more machines doing more for us than we ever imagined possible and we want it at our fingertips like, yesterday.
I remember back in 2010 when the IEEE approved 40Gb and 100Gb and thinking there’s no way we would ever need to call on these colossal speeds. My broadband connection back then was pathetic by today’s standards - like trying to blow a medicine ball with drinking straw - but at the time it served me well because I wasn’t consuming anywhere near the amount of bandwidth that I can’t live without today. But once my kids started using their cell phones, tablets, and online gaming platforms, my poor little copper connection started to groan with the demands that were placed on it. Thankfully today I have fiber delivering over 30x the speed I used to have, but with more and more devices at home, I could always do with more. Much, much more!
Let me give a little bit of background on the technology roadmap to bring us right up to speed (ahem) on where we are today. Up until 2010, life was simple. We were running 10Gb’s across our fiber networks using 1 fiber to transmit and 1 fiber to receive. It didn’t matter whose active devices hung off the end, or which transceiver was deployed, because they all operated the same. And then IEEE ratified 40Gb using a different PMD (Physical Media Dependent) interface with Parallel Optics.
Unlike serial (duplex) transmission, parallel optics uses 8 fibers or more instead of the traditional 2 fibers. We transmit 10Gb per fiber over 4 fibers and receive 10Gb per fiber over the other 4 fibers to reach our aggregated 40Gb speed (known as 40G- SR4). And because the fastest switch ASICs (application-specific integrated circuits) available in 2010 were 10Gb, to reach 100Gb (known as 100GBASE- SR10), also ratified at the same time, we needed to have 20 fibers, 10 each to transmit 10Gb and the other 10 to receive 10Gb. Thankfully in 2015 the IEEE ratified 100G- SR4 transmitting 25Gb over 4 fibers and receiving 25Gb over the other 4 fibers to reach our aggregated 100Gb speed using the same 8 fiber platform as 40GBASE-SR4, and this consistency of fiber counts and electrical lanes continues as we move towards 200G and ultimately the 400G we saw announced this year.
So now this always-on industry presented network managers and facility owners with a choice: As more and more options are added, critical decisions need to be made around what infrastructure solution you should deploy to ensure seamless migration path to higher speeds. Let me break them down for you.
Three years after IEEE approved 40 & 100Gb, late 2013 saw the introduction of 40G BiDi from Cisco using a technology developed by Avago Technologies, now part of Foxconn Interconnect Technologies (a business unit of FOIT). This pioneering technology was developed to enable the existing two MMF fiber networks that were installed in data center facilities using WDM (Wavelength Division Multiplexing) techniques. To achieve 40Gb, two x 20Gb signals on the same multimode fiber but at different wavelengths, or colors of light, transmitting and receiving 20G at 850nm and the other at 910nm ‘bi-directionally,’ hence the name BiDi. We saw the release of 100G BiDi in 2018.
Shortly after the BiDi release, Arista and Juniper launched their Universal transceiver for 40Gb (UNIV) using yet another WDM technique to operate across a single pair of fibers. This time it launched 4 x 10Gb in the traditional single-mode 1310nm region, with center wavelengths of 1271, 1291, 1311 and 1331nm unidirectionally yet still across a pair of MM fibers or SM fibers (hence the universal name). This could also be used across SM fibers as well.
To round out our 3rd WDM alternative, we look at SWDM4. Using a similar 4-wavelength approach to UNIV, the key difference here is that they launch in the more traditional MM 850nm region with center wavelengths of 850, 880, 910 and 940nm with options for both 40Gb and 100Gb.
There are WDM options for 200Gb as well, but the key takeaway at this juncture is that unlike Parallel Optics which is a fully ratified IEEE standard, these technologies remain proprietary or as part of an MSA (Multi-Source Agreement) today and are not interoperable with each other or any standards’ approved transceiver type. BiDi has to connect to another BiDi, UNIV to UNIV and so on. This has the potential to severely limit the options available to network managers as they look to migrate to the next speed, often ‘vendor-locking’ them to a single manufacturer and limiting those critical infrastructure choices I referenced earlier.
It’s very clearly documented that parallel optics, using lower-cost components, are always released into the market first, with their WDM counterparts following 3-5 years later. If you decide to follow a 2 fiber network strategy with WDM and your business/clients demand the next technology upgrade, what choice are you left with if it’s not yet available?
Either stall until it’s released, but this has revenue and reputation impacts - or upgrade the network to support a parallel optic deployment. If you have the fibers in place today, however (even if you are still following the WDM path), you have the option to use both technologies across the same cabling plant without any unplanned upgrades.
Traditional NRZ modulation (Non-Return to Zero) uses 2 values of 0 & 1. Think of it like a lightbulb being the transmitter and your eyes being the receiver. You simply turn the light, or optical signal ‘on and off,’ but if your eyes are as bad as my old things this becomes more difficult to identify or receive as the speeds increase. For 100G, PAM4 (Pulse Amplitude Modulation) was developed which doubled the symbol rate of NRZ and again to simplify the methodology, it uses 4 discrete values 00, 01, 10 & 11 - turning the light slightly on, fully on, slightly off and fully off. PAM4 is used on some, but not all, 100G options - it will, however, become the norm for 400G. It’s worth noting that, as speeds increase, we may see the introduction of other modulation techniques down the line.
Now to keep this article shorter than a bestselling novel, I’ve only referenced Ethernet and the different MM options. Please be aware that single-mode (SM) is by no means any easier. For as long as I’ve walked on this Earth, SM has always been 2 fibers just like our traditional MM brethren until the advent of 40G. And just like MM, we started to see Parallel Optic transceiver options for SM from 40G and above as users needed longer distance, but still wanted to take advantage of port breakout capabilities - taking a larger speed port and breaking it into 4 smaller speeds to leverage a much better total cost of ownership. The shorter distance of 500m PSM4 vs traditional SM distances of 10Km results in a lower cost.
The current industry trend is that once the volumes start to take hold, the cost per port to deploy a higher speed parallel optic and break it out, versus using the same native speed optics (1 x 100G broken into 4 x 25G rather than 4 x native 25G as an example), becomes lower and the higher the optic speed, the more cost-effective it becomes. In addition, breakout offers significant density savings as well as reduced energy costs for both power and cooling.
400G - where we are today
As I mentioned earlier, our insatiable appetite for bandwidth has driven the industry to develop even faster, more latency-independent networks. The proliferation of Artificial Intelligence (AI) & Machine Learning (ML), serverless compute, distributed flash storage and faster server CPU/GPU/FPGA are all factors driving 400G. As an example, Broadcom’s Tomahawk III ASIC supports 12.8Tb/s, or 128 x 100G optical signals, and there are developments for 25.6Tb/s (256 x 100G ), 51.2Tb/s (512 x 100G ) and 102.4Tb/s (256 x 400G ). Compressing more and more throughput into the same footprint provides the opportunity to process data much faster than ever before, fulfilling the ‘Need for Speed’ at our fingertips. This doesn’t come without its challenges and you’ll see in the table below that we are being presented with several different options to support length, lane and fiber types.
Here is a list of all the current 400G optic PMD’s (Physical Media Dependent) that have either been released or are still in development as an IEEE standard:
In addition to the above, 2016 saw the formation of two MSAs looking at transceiver form factors for 400G with eight electrical lane connections. The form factor MSAs were developed to look at what the connectivity options would be, including the packaging for the transceiver. These are QSFP-DD (Quad Small Form-Factor Pluggable – Double Density) and OSFP (Octal Small Form-Factor Pluggable). Corning is a member of both these MSAs.
QSFP-DD - http://www.qsfp-dd.com/
QSFP-DD uses the same physical footprint as QSFP28, so it’s fully backwards compatible when using the same chassis or line card. In terms of optical interfaces, the most recent specification (Rev 5.0 July 2019) lists the following options :
- Male MPO
- MPO-12 Single Row
- MPO-16 Single Row
- MPO-12 Two Row
- Duplex LC
- Dual CS
*MDC and SN are a new generation of Very Small Form-Factor connectors that were launched early in 2019, to be covered in a separate article. These are being driven by Hyperscale operators who would like to break out 400G to 4x100G optical circuits directly at the transceiver.
OSFP - https://osfpmsa.org/.
The OSFP form factor was designed with higher speeds in mind (however both QSFP-DD and OSFP can handle 800G heat dissipation), and as such has a slightly larger footprint than the QSFP-DD. This is mainly due to an additional set of cooling fins along the top of the body designed to dissipate higher temperatures generated as a result of the higher power required (note that QSFP-DD has added to the design to dissipate more heat now). These temperature increases are consistent for both QSFP-DD and OSFP with a typical requirement for multimode between 8-15W. To offer backwards compatibility to QSFP28, you need to use an adapter.
The current OSFP specification from January 2019 lists :
- Male MPO
- MPO-12 Single Row
- MPO-16 Single Row
- MPO-12 Two Row
- Duplex LC
- Dual CS
The more astute amongst you may have noticed that the table of PMDs does not align with the options listed within either MSA, and you would be completely accurate. The reality of the matter is that while the different optical interfaces are listed inside the MSA, there are currently only identified optics for CS, MDC or SN, and no identified solution for MM LC, just SM for 500m, 2 & 10km.
“So when will we see 400G optics coming to a store near us soon?”, I hear you all cry! We saw announcements from Arista and Cisco early this year that expect to have product ready for shipping by the end of this year or early into 2020. Arista have announced options for both QSFP-DD and OSFP, and Cisco just for QSFP-DD.
To Infinity and Beyond, or rather 400G, may seem quite a way off for many of you reading this article. What we’ve all come to realize, however, is that we as an IT industry will never stop pushing the envelope and we must all accept that what we love today will soon become the memory of yesterday.
We have experienced more network speed increases in the last nine years than we have in the previous 30 (with more to follow). With so many options, planning the right technology migration path with the most flexible and scalable fiber infrastructure possible should be at the top of your priority list. If you don’t, you could find yourself having to continue investing in additional fiber, because the optics to support your next technology refresh are not yet in development.
You can’t afford to be caught napping because to quote the Toy Story toys: “We are everywhere!”