100G Ethernet

Arista's On-Board Optics Boost 100G Density

The newest Arista Networks Inc. router packs what appears to be the industry's densest 100Gbit/s so far, a feat accomplished by putting the optics directly onto the boards.

The 7500E, announced Wednesday morning, puts 12 100Gbit/s ports on each card, for a total of 96 per chassis.

Silicon photonics aren't involved. Arista glommed vertical-cavity surface emitting lasers (VCSELs) directly onto the board, obviating the need for transceiver modules. End users plug fibers directly into the box without having to buy optics.

Remember that Arista founder Andy Bechtolsheim helped spread silicon photonics fever a couple of months ago during Facebook's Open Compute Summit. He shared the stage with Intel Corp. and advocated the wonders of the technology.

But silicon photonics is "not ready for prime time," so the technology didn't make the cut for the 7500E, says Martin Hull, an Arista senior product manager.

Still, by packing a transceiver's guts directly onto line cards, Arista gained a lot in density. Using CFP transceivers, it's possible to get four 100Gbit/s interfaces onto a card. The newer CFP2s, which have barely begun shipping, can fit eight. So, jumping to 12 interfaces per card is pretty significant.

Each of Arista's 100Gbit/s ports is actually 10 lanes of 12Gbit/s apiece, so we're not talking about coherent transmission for long distances. It's a data-center switch, using SR-10 optical standards for a reach of 150 meters. But the setup does let operators change the port speed in software, dividing it into groups of 10Gbit/s or 40Gbit/s connections as necessary.

Optics represent a relatively high percentage of a switch or router's cost; Arista estimates it at 20 to 25 percent. A 10Gbit/s SFP module would cost $450, by Arista's estimate.

Putting the VCSELs onto the board incurs a cost, too, but Arista is claiming its interfaces are a clearly cheaper alternative.

According to Anshul Sadana, Arista's senior vice president of customer engineering, a typical switch, when fully populated with 100Gbit/s interfaces, will cost $35,000 to $50,000 per 100Gbit/s port -- that is, dividing the cost of the entire switch by the number of ports. A fully loaded Arista 7500E would come out to $10,000 per 100Gbit/s port.

Arista claims to have kept the power nearly constant as well. That is, the Arista 7500 uses about 10 W per 10Gbit/s port, while the 7500E, which can support three times as many ports, uses 4 W per 10Gbit/s port.

Lower price-per-port was the trick that made the Catalyst 6500, Cisco Systems Inc.'s flagship switch, thrive after its release in 2003, writes ISI Group Inc. analyst Brian Marshall in a note published Wednesday. An Arista fan for some time, Marshall sees the 7500E having a similar potential, possibly helping Arista float an IPO "in the next four to six quarters."

Then again, Cisco has its own optical projects in the works. In addition to producing the CPAK, its own CFP2-like transceiver, Cisco is applying silicon photonics to chip interconnects. It's not a stretch to think an on-board transceiver could be in Cisco's grasp as well. (See Cisco Goes Inside With Silicon Photonics.)

The 7500E is the same chassis as the Arista 7500 switch; the difference is in the cards, including a new fabric card that triples the fabric bandwidth, to a theoretical 30Tbit/s.

The 7500E is in general availability, with some customers already running it in production environments, Arista says.

— Craig Matsumoto, Managing Editor, Light Reading

Jeff_k_M 5/6/2013 | 11:44:03 PM
re: Arista's On-Board Optics Boost 100G Density It could be true that no one has officially passed Telcordia, but this is what I was eluding to in my first post. If you use these in a decent environment and don't mechanically stress them too much, they perform quite well. They're raw optical performance isn't far different than that of a single fiber connector, but yes when you stress them they don't perform as well. There's has to be some trade-off when you cram many fibers in just about the same space, as a single fiber connector. A big part of the problem with these connectors is that, even after all the years they've been around, still so many people do not understand how to polish them correctly and perhaps just as important, clean them properly after polishing.
Jeff_k_M 5/6/2013 | 11:39:43 PM
re: Arista's On-Board Optics Boost 100G Density In GR-1435, for media type 2 proof testing you bend the fiber perpendicular to the connector, but in the same plane. Meaning you only bend the ribbon the way it was designed to be bent. So it won't matter if it's too stiff in the one direction. What is all this about only using the middle 8? I have made, tested, and used at least a thousand 12 fiber MTPs, with all 12 fibers in the ferrule. I didn't see any issue? Is this an OSP thing?
redface 5/4/2013 | 6:20:46 AM
re: Arista's On-Board Optics Boost 100G Density MT connectors have been a necessary evil. They have the form factor required, but the performance is far inferior to that of single channel connectors. I think MT connectors have never passed the Telcordia reliability tests. This is the reason that MT connectors are not as widely deployed as their functionality would suggest.

Even the ribbon fibers are problematic. Which is why you only use the inner 8 fibers of a 12 fiber ribbon. The ribbons are too stiff in one direction...
cw.774 5/3/2013 | 10:02:12 PM
re: Arista's On-Board Optics Boost 100G Density yes - It is the 8 inner fiber positions. This avoids the problems with using the end positions. I wonder why you think it's a reliability concern?

My understanding has been for the last 10 years that MT should not be a reliability risk here. It is a stable technology proven for single mode applications so these multimode applications shouldn't be a risk. It's little brother (mini MT) is a problem however and proved useless for single mode fiber,
highfiber 5/2/2013 | 7:36:38 PM
re: Arista's On-Board Optics Boost 100G Density If people would accept the fact that you must clean fiber end faces before plugging them in, there would be a lot less unreliability. The fact that you don't clean 12 fibers at a time in an MTP just means you have a higher probability of an issue with the whole MTP interface then when you are only not cleaning two fibers in an LC connector.
Jeff_k_M 5/2/2013 | 2:21:19 AM
re: Arista's On-Board Optics Boost 100G Density Short of being used in an OSP environment there is nothing terrible about an MTPs reliability. Assuming you can withstand the slightly higher IL in your budget, things should be fine.
redface 5/2/2013 | 12:48:41 AM
re: Arista's On-Board Optics Boost 100G Density Might be a little more complicated than you think

For VCSELs, there is no wavelength combining. Each 12Gb/s resides in a separate fiber. So probably each 100G port uses one 12 channel MTP array connector (the outermost fibers are probably unused). For a card with 12 ports, 12 MTP connectors will have to be used. However, MTP connectors have far inferior reliability compared to single channel connectors. I think there might not be insignificant problems with the Arista design.
Sign In