SC13 – Updated The current top-end Ethernet standard may be 100 gigabits per second, but don't expect the next step up to be one terabit per second. The days of Ethernet speeds improving by an order of magnitude are gone. Why? Because nobody wants to pay for the necessary research and development – nobody in the US or the EU, that is.
"My view is that 10X Ethernet increases are not going to happen anymore," said Chris Cole, engineering director at Finisar, but representing just himself during a discussion session at the SC13 supercomputing conference in Denver, Colorado, on Thursday.
"One of the big reason that you had 10X increases was that the telecom guys paid for the technology development," Cole said. "Those were the days when AT&T and the carriers in other countries had big checkbooks and they could fund things like Bell Labs. So they paid for some very advanced technology which then allowed Ethernet catch on. Those days are over."
The new advancement rate will likely be 4X, not only for Ethernet, but also for the core optical transport rates set by the ITU-T. Some industry types are still holding out for a 10X jump to terabit Ethernet, but odds are heavily stacked that the next big jump in Ethernet speeds will be to 400Gbs – aka 400 gigabit Ethernet, or 400GbE.
Just last week, in fact, the IEEE's "400 Gb/s Ethernet Study Group" approved objectives [PDF] for four different link distances of 400GbE, beginning the long and complex process that will eventually result in a full, certified specification, probably around 2017.
The IEEE project number for the 400GbE effort is – no tittering, please – 802.3bs.
Cole doesn't believe that terabit Ethernet is in the cards. "My crystal ball says that the next Ethernet rate after 400G will be 1.6 terabits." Arguing against the possibility of a jump from 400GbE to 1TbE, he doesn't see the economics working in 1TbE's favor.
"That's only a 2.5X increase," he said, "but everyone agrees that to get to a terabit will require completely new technologies, and 2.5X just doesn't feel like a bit enough step to justify a huge R&D investment."
There's that almighty dollar again.
Nobody wants to pony up the hefty amount of cash required for the move to 1.6TbE. As Cole noted, the telecoms no longer want to pay for R&D – they want the suppliers to front the money.
"Of course the supplier base has no money," he said, "so we keep searching for the guys with money in their pocket." He then told his audience of HPC pros that "at a meeting of the IEEE last week, there was no one from the HPC community there, and so there were a number of suggestions that maybe HPC should pay for the next generation of technology – at least in some quarters you're perceived as being wealthy."
Interestingly – and somewhat worryingly – Cole told The Reg that the R&D funding problem is most severe in the US and EU. "In Asia you see the large system OEMs funding R&D extensively. In the US the suppliers actually want to fund it, they just don't have the money."
From his point of view, the market is still in bad enough shape that customers have sufficient leverage over suppliers to depress prices. As a result, margins are squeezed to such a degree that at the end of the day there's not enough cash left over for suppliers to fund R&D.
China, on the other hand, is shoveling money into R&D, he said. "It's good news, bad news. The good news is there is some of the advanced work being done, and the bad news is that the jobs will probably shift over time to Asia."
By then, those Asian researchers will likely be able to transfer their data sets back to their fellows in the US and EU over 400GbE links – if, of course, there are any Ethernet engineers left outside Asia. ®
In the original version of this article, Chris Cole's statements were inadvertently attributed to a different member of the discussion panel.