400, and even 800Gbit/s Ethernet standards en route
SAN JOSE, Calif. — As many as a dozen companies will demo at this week’s Optical Fiber Conference (OFC) prototype 400 Gbit/second Ethernet optical modules that data center and telco networks expect to deploy starting late next year. Meanwhile, standards efforts are already plowing a path for the next generation to keep up with rising demands.
The emerging 400G products typically use eight 50G serial connections based on the latest PAM-4 modulation techniques. Last week, the IEEE gave preliminary approval to start work on standards for 100G serial links that could drive future 800G Ethernet standards.
Meanwhile, the Consortium for On-Board Optics (COBO), an ad hoc group of about 70 companies, released its first specification for modules embedded on routers and switch motherboards. The spec aims to reduce heat and power issues for 400-800G products as the first step toward a future of merged optical and electrical components.
All the efforts face tough balancing acts achieving the highest possible data rates while maintaining acceptable distance, power, heat and cost. They come at a time when optical links are now widely used inside data centers and carrier central offices.
Today’s large data centers are transitioning from 10 to 25G Ethernet inside racks of servers and deploying 100G Ethernet between the racks.
“400G is the next phase…People are hoping by late 2019 or early 2020 to start deploying, and it looks like a lot of the components are coming on line now — we’re already seeing the switch silicon showing up,” said Brad Both, a networking specialist for Microsoft’s data center group and chair of COBO.
He expects to meet with as many as a dozen optical module vendors showing their first 400G products at the event. AT&T Labs will give a paper at OFC on its field trials of 400G Ethernet, and engineers from eight companies including Google and Microsoft will discuss 400G optics in an OFC panel
COBO drives optics to the motherboard
COBO’s approach takes optics off a module that users plug into the faceplate of a router or switch and embeds them on the motherboard. It enables better airflow and larger heat sinks to reduce thermal requirements for systems as data rates rise.
The first generation 400G modules emerging at this year’s OFC are expected to consume 12-14W. COBO defines 8- and 16-lane versions that support 20 and 40W, respectively. The group will release a half-dozen white papers in the next few weeks detailing implementation issues.
COBO expects initial demo products this fall with most implementations aiming at whatever becomes the next speed grade after 400G. It will start work soon on a follow-on standard targeting whatever comes after 800G.
“Eventually optics must move into the switch ASIC because serdes already consume 30-40 percent of a networking chip’s power,” said Booth, citing longstanding debates about when and how the merger may happen.
With COBO, OEMs will configure their systems with optical modules and test them before the ship. Today data centers and carriers buy optical modules themselves and install them in the field.
It’s a big change, but one Booth said is like what happened with servers that are built up from multiple PCI cards and memory modules and shipped to users. “Networking is still five years behind the server,” said Booth, who formerly worked at Intel.
COBO is an alternative to an alphabet soup of options for pluggable optical modules based on ad hoc standards called multi-source agreements. They include the QSFP-double density approach backed by Molex and Facebook, and OSFP with support from Arista and Google.
“There are too many choices…Thankfully, we’re 18-24 months before deployment, so we can start evaluating module form factors…people may work on the thermally hardest ones first,” said Booth.
A variety of other standard and proprietary approaches for 400G covering different distances are still in the works, said Ben Rubovitch, CEO of DustPhotonics, a module maker that aims to use OFC to explore what its customers want.
The startup will show prototype 400G modules it will sample soon, supporting 100-meter reach at the event. It expects to support multiple form factors and see approaches for multiple kilometers in the not-too-distant future.
A look down the road from 50-800G+
“Today we see plenty of big 400G customers that have systems in their labs… some are trying to ship by the end of this year,” said Rubovitch, a former Intel and Mellanox engineer that brought DustPhotonics out of stealth mode at OFC.
The deployments are not expected to start in earnest until 2019, when perhaps half a million 400G modules will ship. That could ramp to 3.5 million units a year later with roughly equal thirds using copper, multimode and single mode fiber, he said.
DustPhotonics hopes it can break even by June of next year in part with help from sales of a 100 and 200G modules it designed to prove its optical technology. The startup got its first seed money just 15 months ago and closed a series A for an undisclosed amount from private investors including Lip-Bu Tan of Walden International.
“They are aggressive, and I will take a look at what they have…but they will not be the only one at 400G,” said Booth of Microsoft which has already put out a request for information on 400G modules.
“There are a lot of people stepping up to 400G, including a number of startups…and some who missed the shift to 100G…This sector is ripe for growth and acquisitions in the next two years,” he added.
After a short break, COBO will turn its attention to what may come at 800G and beyond. The Optical Internetworking Forum is already working on 100G serial standards that could be the basis for a future 800G Ethernet spec, however the OIF’s work is so far focused on short reach links.
The good news is the 100G lanes are expected to leverage the existing PAM-4 modulation that today’s 50G serial links are pioneering. The Ethernet Alliance (EA) is expected to set up an interoperability program soon for the emerging 50G products, many of which will be shown at OFC by vendors such as Marvell and Xilinx.
With 50G the industry shifted away from the simpler NRZ modulation to PAM4. Thus, “there’s still a lot of work to get done, tested out and mainstreamed,” said John D’Ambrosia, who leads the EA group as well as many IEEE Ethernet initiatives.
The IEEE is still early in its process of organizing a 100G serial standard effort, said D’Ambrosia, noting it could take three years to finish a spec. Eight of the lanes would be an optimal way to create an 800G Ethernet standard, he said, though some may choose to try 16 50G lanes.
“I’m more comfortable with 8x100G, and 16 lambdas at 800G scares the heck out of me,” he said, citing signal integrity and other issues.
“As we go faster, some solutions don’t go as far so we have to make a leap of some kind…some propose coherent solutions and we need to ask a lot of questions about how far we really need to reach,” he added.
— Rick Merritt, Silicon Valley Bureau Chief, EE Times