Arista Champions New 100GbE Optical Spec, Files To Go Public
Frustrated by the limitations of the current crop of transceivers that have been approved by the IEEE and adopted by the industry, Arista Networks is teaming up with Intel and two dozen customers and parts suppliers to create a new transceiver specification that will suit the needs of hyperscale datacenters.
The news comes hot on the heels of Arista filing its paperwork to do an initial public offering of its stock on the New York Stock Exchange.
The new specification will be created by the CLR4 100G Alliance, which is being run by Arista and Intel, and according to Andy Bechtolsheim, the chief development officer at Arista, bridges a gap that has been left open by the current collection of transceivers and, importantly for Arista, will limit the adoption of 100 Gb/sec switching in datacenters.
"When the IEEE created the 100G standard back in 2010, there were two optics in that spec. There was the 100G SR10 and 100G LR4. Unfortunately, neither one of those addressed the needs of the large datacenter operators, which need a reach of multiple hundreds of meters, up to a kilometer or two, and most importantly needed a power-efficient, cost-effective design that the IEEE LR4 standard is simply not meeting." In the past several years, explained Bechtolsheim on a call about the new alliance and the standard it is proposing, there were a number of IEEE efforts to try to come up with next-generation optics for 100 Gb/sec Ethernet switches and routers. "Unfortunately, none of the cost-reduction proposals for optics relative to the datacenter made the vote, and as a result, the objective to create a lower cost, lower power standard for 100G in the IEEE failed and was abandoned last summer. The problem in the market has not changed, meaning it is now more important than ever that we have a low cost solution for 100G in the datacenter."
The 100G CLR4 standard being put forth by Arista and Intel is designed to span ranges of between 100 meters and 2 kilometers. This range is important. As an example, Bechtolsheim showed an aerial view of Microsoft's datacenter in Quincy, Washington, which weighs in at 800,000 square feet and which is 300 meters across. To go from end to end, as Microsoft needs to do in that datacenter to allow for so-called east-west traffic between clusters arranged around the datacenter, would require the software and now cloud giant to use very expensive switches and optics designed for carriers. This is not practical for economic and technical reasons. Bechtolsheim said that in hyperscale datacenters, the switch-to-switch links for backbones were often 500 meters, often 1 kilometer and now in some cases 2 kilometers apart.
The 100G CLR4 spec will use a transceiver that employs coarse wavelength division multiplexing (CWDM), running four wavelengths of light across two single-mode fibers to provide duplex networking. Importantly for Arista, the 100G CLR4 spec will use the existing QSFP form factor for the transceivers, which has already been used for 40 Gb/sec uplinks and switches. The goal is to have a port that consumes less than 3.5 watts and that allows for up to 36 ports running at 100 Gb/sec speeds in a 1U chassis.
As EnterpriseTech reported last week, Arista has just rolled out a set of 100 Gb/sec line cards for its 7500E modular switches using MXP, CFP2, and QSFP ports. Arista can cram a dozen QSFP ports running at 100 Gb/sec onto the line card, but only six CFP2 ports.
Bechtolsheim said the plethora of transceiver form factors being developed for 100 Gb/sec networking was confusing the market, and that the number of different options exceeds those that were available for 10 Gb/sec and 40 Gb/sec switches. At the moment, there are six different form factors, with five of them (excluding CFP4) shown below:
Bechtolsheim called this "surreal" and said it was even more perplexing that there are six different transceiver types in a market that does not yet even ship more than 10,000 ports per quarter. This situation confuses both switch, cable, and transceiver makers, who are not sure what they need to build. "The important thing is to get to high density and low power, and the only thing that really makes sense is QSFP," he said.
The alliance formed by Arista and Intel, which supplies its Fulcrum switch ASICs to Arista and other switch makers, is open to switch, transceiver, and system manufacturers as well as end user customers, and 20 companies in total had joined the effort as it was being developed. eBay is the big marquee customer behind the effort, but with Microsoft running some of the largest datacenters in the world and, as it turns out, being Arista's largest customer, you can bet Microsoft is watching the effort very closely and perhaps participating. Hewlett-Packard, Dell, Fujitsu, Brocade Communications, and Oracle have signed on to the alliance as well.
The 100G CLR4 specification will be available at no charge, and a preliminary version is available now to members. The idea is to get feedback and get a finished spec out the door sometime in May.
"Between last July and now, nine months have passed," said Bechtolsheim, with a sense of urgency. "The key thing here is that we cannot wait another nine months for a spec to come to fruition. Arista announced its first QSFP line card last week, which we will be showing at Interop, and we need those optics now."
In a separate development, Arista has filed its <a target="new" href="http://www.sec.gov/Archives/edgar/data/1596532/000119312514122171/d639957ds1.htm">Form S-1</a> with the US Securities and Exchange Commission, and from this we can get a glimpse into the hyperscale datacenter world where the company has been operating.
In that document, Arista cites statistics that show the large Internet companies in the United States have boosted their capital spending from $8.9 billion in 2010 to $19.4 billion in 2013, a 30 percent compound annual growth rate. The company also cites datacenter switching revenue figures from Crehan Research, which shows the 10 Gb/sec Ethernet switch market (not including switches embedded in blades) as growing from around $6 billion in 2013 to $12 billion in 2017, which is a 19 percent compound annual growth rate. This is the portion of the market where Arista got its start and has done well, selling over 1 million ports and amassing a customer base of 2,340 companies in the six years it has been selling products.
Here is a summary of the important financial figures that Arista revealed in its S-1 filing:
The company has been growing fast, as you can see, but revenue growth stalled in 2012 for some that Arista did not explain. The interesting bit is how dependent Arista is on those hyperscale datacenter operators. Arista counts six of the largest cloud operators as its customers, including Facebook and Yahoo as well as Microsoft (it did not name the other three, but Amazon Web Services and Google are probably on that list). It also called out Comcast, Equinix, ESPN, and Rackspace Hosting as large customers – EnterpriseTech just profiled the 100 Gb/sec rollout at Equinix last week – and in the financial services sector where Arista got its start with low latency 10 Gb/sec switches, Barclays, Citigroup, and Morgan Stanley are big accounts.
The top ten customers accounted for 43 percent of sales in 2013 at Arista, and Microsoft was more than half of that, as you can see in the table above. So when Bechtolsheim says he needs cheap optics that have range now, he is not kidding and that is why he has teamed up with Intel and others to try to get something into the market that meets the needs of these customers.
The other interesting thing is the amount of money Arista spends on research and development. The company does not design its own ASICs – it relies on chips from Intel and Broadcom solely and has from day one – but plows most of its dough into software development to make its Linux-based Extensible Operating System for its switches as malleable as possible.
The thing that Wall Street will like is that Arista actually makes money. Its margins are on par with much larger rival Cisco Systems, in fact, and that is a bit of an accomplishment in the cut-throat switching space.