SAN JOSE, Calif. — A standard to slash the power consumption on Ethernet networks is gaining traction and could be in a draft form suitable for designing products as early as September. The so-called energy-efficient Ethernet group has already settled on an approach for use on 100 Mbit and Gbit chips.
"I hope we can come out of our meeting in September with [a draft] people could implement," said Mike Bennett, a researcher at Lawrence Berkeley National Lab chairing the effort officially known as IEEE 802.3az.
The group has already accepted a so-called low-power idle proposal for use with 100 Mbit and Gbit connections. Under the proposal, led by Intel, Ethernet chips with no data to send would be able to put the PHY into a sleep mode.
Once the PHY is in sleep mode, engineers could decide whether they wanted to put other parts of their systems into a low power state. "I think a competitive edge could come from how people turn things off in a system based on what's happening at the PHY," said Bennett.
However, at 10 Gbit rates, some engineers have suggested the low-power idle approach could require additional buffers or might introduce jitter or latency that would be unacceptable. An alternate proposal from startup Aquantia and Broadcom, called subset PHY, calls for scaling a 10Gbit link down to a single Gbit channel to save power.
Both proposals will require some changes in the PHY and MAC components which vendors would probably make during regular product refresh cycles. "There are not huge hardware changes, but there are some," said Robert Hayes, an Intel networking engineer who helped define the low-power idle proposal.
A Broadcom executive said in a recent interview that the subset PHY approach is better suited to networking equipment. The low-power idle approach is geared more closely to the interests of computer servers and clients, he added.
However, Hayes said engineers from Hewlett-Packard's networking group as well as Cisco Systems have shown support for low-power idle. Unlike the subset PHY approach which maintains a Gbit connection, it allows engineers to turn off a range of functions in their systems, he said.
"If we know we don't have data coming down the wire we can turn off a PCI Express bus, memory controller and maybe even parts of a host processor," said Hayes. "That leads to substantial savings in a server and some switch vendors are saying the same thing," he added.
The .3az committee needs to debate the merits of the two approaches at meetings in May or July. It is unlikely any new proposals will come up at next week's meeting, the last meeting where new technology ideas can be proposed.
The group is also looking at proposals for 10 Mbit and backplane Ethernet.
The group is likely to propose lowering the transmit amplitude for 10 Mbit Ethernet signaling originally designed for the Category 3 cabling of its time. Today's Category 5 cables can handle lower transmit amplitudes, saving power.
Separately the group is looking at ways to apply low-power idle or subset PHY techniques to Gbit and 10 Gbit backplane Ethernet links. The trade offs for backplanes between the two techniques are expected to be debated at the meeting next week.
Bennett said he was hopeful a decision on all the outstanding issues could be reached by the end of the September meeting.
The .3az group held discussions with representatives of the 802.1 committee on higher player protocols. The .3az goal is to create mechanisms that would let higher layer protocols override energy saving techniques in the PHY if an application or operating system needed to choose performance over low power.
The group initially explored the idea of ratcheting speeds down to as low as 10 Mbits/s when there was no network demand. However, tests indicated the switching time of as much as 20 milliseconds from low to high speed could cause unacceptable packet losses, especially for audio traffic.
The .3az work is in part a spin out of a broader initiative at Lawrence Berkeley Lab to lower energy consumption across a wide range of network and consumer systems.
One Lawrence Berkeley Labs researcher estimated the IEEE effort alone could save $450 million in energy costs a year just in the U.S. Interestingly, the brunt of the savings--$200 million--could come from home computers, with another $170 million from offices and $80 million from data centers. The researcher estimated that a Gigabit interface card could shave up to 1.5 W off its power consumption and a 10-Gbit card up to 10 W by scaling back data rates at times of low traffic.
