Forget the predictions that copper-based 10 Gigabit Ethernet would run on 10Gbase-T. Copper-based Ethernet is here, but SFP Direct Attach maybe the cabling of choice.
Bryan Betts
Four years ago, the networking industry was gearing up for massive uptake of copper-based 10 Gigabit Ethernet (GbE) using 10GBase-T, but expectations were never fulfilled. Most network managers didn't want to invest in expensive 10GBase-T and most 10 GbE adoption was in backbones and telco-type applications, typically over optical fibre.
Yet now that 10GbE is booming with new applications accentuating the differences between Ethernet adapter designs, copper is back on the 10GbE menu. Only now the copper option is in a newer format called SFP Direct Attach (SFPDA), which uses twin-axial copper cable and plugs into the SFP+ receptacles also used for fibre optics.
"We have again raised our forecast for the 10G Ethernet controller and adapter market, as shipments during the past two quarters have come in higher than we had predicted," says Tam Dell'Oro, president of market research company Dell'Oro Group. She forecasts that 10G port shipments will increase to just over 34 million in 2015 from 3.1 million in 2010, a 61% CAGR.
Melvyn Wray, senior VP of product marketing EMEA at switch maker Allied Telesis, says that while the range enabled by 10GbE Direct Attach is limited – nominally 15m, but nearer to 5-10m in practice – it is ideal for top-of-rack or cluster switches. It allows an SFP+ device to use either copper or fibre, it is low power, and its cabling is relatively thin and light.
"10G has become almost mandatory in the data centre now," Wray says. "For the last few years the choice has been CX4 [copper] or optical fibre, but CX4 is a big bulky cable and unfortunately optical prices have been quite high -- about $1000 for SFP+ modules and fibre. 10GBase-T transceivers draw about 4W, so that's 8W per connection – that's a hell of a lot of power just to transfer a signal maybe a few metres.
"Over the last nine months or so, twin-ax copper with an SFP+ module at each end has come down to around 0.2W and a maximum of $200 per link. It's very flexible, low cost, easy to manufacture, and has near-zero latency. There's no need for optical conversions." This price-point means that a 10GbE link is finally around the same cost as the alternative approach of four or five 1GbE links bonded together, he adds.
Will there be a future for 10Gbase-T in copper-based 10 Gigabit Ethernet?
Those with long memories might hear echoes of the original 10Mbps Ethernet here. That started off on thick coax cable, which, like CX4, was stiff and hard to route, and it only really took off once an easily-handled thin coax version appeared. The question today for 10GbE is whether there will be an even more successful third generation based on twisted pair copper, as there was with 10Base-T.
Steve Pope, CTO of 10GbE controller and adapter developer Solarflare there will be. 10GBase-T technology is still evolving, he says, but in a couple of years adapters will be nearer 1W power consumption. Then the cheaper cabling and longer range of twisted pair will come to the fore -- at least, outside areas such as finance, which need the lower latency of Direct Attach or fibre.
Popes, however, that there is one notable way in which the 10GbE market will be unlike earlier generations: he expects to see relatively little embedded LAN-on-motherboard (LOM), with users preferring adapter cards and mezzanine cards even though they're more expensive.
"Gigabit Ethernet was completely commoditised with zero value-add, but with 10G the network controller business is diversifying again and there are now real markets where customers want differentiated solutions," Pope says. Different chipsets and software stacks really do perform better in some applications than others - iSCSI, FCoE, high-performance clustering, and so on -- so server suppliers need to offer options.
"The different controllers are genuinely different, so OEMs can't build a single 10G platform the way they could for 1G," Pope says.
That certainly ties in with Dell'Oro's observations: “What really took us by surprise was the strength in network adapters rather than LOM,” she says.
“Intel and Solarflare just about doubled their adapter shipments quarter-over-quarter, while Emulex’s were up over 60 percent, and all this is on the older generation servers that support 50G of in/out throughput. What do you suppose will happen when the next generation servers with over 100G of throughput start shipping?”
0 Comments