100-GigE Takes Shape

SAN JOSE, Calif. — Even though it's going to be three years before a 100-Gbit/s Ethernet standard can be ratified, some researchers and CTOs are already looking at how such a speed could be made possible, with equipment that would be practical enough to actually get sold.

They convened earlier this week, spending a day at the historic Dolce Hayes Mansion -- now a conference center -- buried in the suburbs southeast of downtown San Jose. There, an engineering-minded audience packed a small meeting room to the gills for an Optoelectronics Industry Development Association (OIDA) seminar entitled, "100 Gbit/s Ethernet: The Next Challenge for Communications Systems."

The topic is getting a lot of publicity lately, particularly with the Institute of Electrical and Electronics Engineers Inc. (IEEE) approving a Higher Speed Study Group to standardize the next speed grade. (See 100-Gig Ethernet Takes First Step.)

That standard, which is expected to take three years to complete, doesn't have to be 100 Gbit/s. The HSSG, which won't meet formally until later this month, will have to entertain comments on several speed-grade proposals, including 40, 80, 120, and 160 Gbit/s. Most bets, though, seem to be on 100 Gbit/s.

The loudest cries for a higher-speed Ethernet come from places like carrier hotels, some of which are already reporting a need for 100-Gbit/s links among large ISP tenants. Link aggregation -- a method of treating multiple 10-Gbit/s links as one -- isn't expected to be efficient enough to handle traffic growth. (See Ready for 100-Gig Ethernet? .)

Given that the need is already surfacing, many speakers said designers should be thinking beyond 100 Gbit/s. Drew Perkins, chief technology officer at Infinera Corp. (Nasdaq: INFN), titled his presentation "TbE [Terabit Ethernet] or Bust."

"Core network links blew through 10 Gbit/s years ago, frankly," Perkins said. "Let's not stop at 100 Gbit/s."

Others agreed that there's some urgency to all this. One speaker argued 100-Gbit/s Ethernet is already late for the high-performance computing industry, at least when compared with the timing of previous technologies like 10-Gbit/s Ethernet. "If it doesn't move quickly, other standards like Infiniband may take its place in that market," said Petar Pepeljugoski, a scientist with IBM Research.

Unlike 10-Gbit/s Ethernet, the 100-Gbit/s variety will have to start life with multiple ports per line card, said Joel Goergen, chief scientist at Force10 Networks Inc. "If, [in 2009], you're not targeting your systems designs for 400 to 500 Gbit/s per blade, its going to be difficult for you to compete," he said.

As for the makeup of a 100-Gbit/s port, the endgame would seem to be serial 100-Gbit/s. "Could it happen? Probably 2013," Goergen said.

Most speakers agreed, although Jack Jewell, chief technical officer of Picolight Corp. , wondered aloud if serial 100-Gbit/s Ethernet, while attractive, would be worth the trouble. "I don't really see a need for a completely serial solution," he said during his presentation.

For near-term designs, the point is moot: Conventional silicon can't handle speeds much higher than 10 Gbit/s, so a 100 Gbit/s feed has to be split up somehow. But what combination works?

Goergen, Perkins, and others seemed to gravitate toward using five 20-Gbit/s lanes packed into a module. That's analogous to what's done with Xenpak 10-Gbit/s Ethernet modules, which produce four channels of 3.125 Gbit/s apiece. In fact, the 5x20 idea is attractive partly because it would fit inside a Xenpak module, Perkins said.

Another possibility is to use 10 lanes of 10-Gbit/s Ethernet, but Goergen worried that this wouldn't allow for the kind of density he thinks systems will need. "I can't get four of those types of ports on a blade," he said. Still, other speakers noted that 10x10 might be viable for some applications.

How about four lanes of 25 Gbit/s each? Goergen said he'd backed that idea in the past but now disdains it, because it's too difficult to get that kind of speed to comfortably run on mass-produceable chips. "I worked with two serdes companies on this, and I can't get it to work," he says.

— Craig Matsumoto, Senior Editor, Light Reading