Achieving low power in the data center with 10GBASE-T

The multi-rate capability and ease of use of 10GBASE-T connections are well known, but not so well known is that it is also a low-power technology, building upon process advancements and device innovations. To demonstrate this, this feature will use a fully populated data-center rack as a platform from which to explore the impact that the choice of interconnect will have on total power dissipation in a real-world environment.

Specifically, it will examine the power consumed by the fully populated rack when deployed with 10GBASE-T, and compare that to an equivalent fully populated rack deployed with a competitive technology, the SFP+ with direct-attach twin-ax copper cabling.

Certainly 10GBASE-T did not enter the market as the lowest-power physical-layer (PHY) technology. The challenge of implementing 10GBASE-T requires that the PHY perform extensive processing. However the combination of on-going process shrinks, along with innovations in implementation, provide for significant reductions in power. Consequently, the industry is moving from current-generation devices in the 6-W range to new-generation devices at the 40-nm process node that are projected to be around 3 W.

RJ45 connectors and twisted pair cabling are commonly deployed at the edge of the network and PCs and servers today use gigabit LAN-on-motherboard (LOM). 10GBASE-T has been defined and developed with this legacy in mind. And now, we can see the fruits of these efforts coming to bear as LOM is beginning its transition to 10GBASE-T.

As used in the data center
The sweet spot for 10GBASE-T is at the edge, at the interconnect between the server and the first-level aggregation switch. Historically, these connections benefited from the flexibility of structured cabling and would find use over a wide distribution of cable lengths, from sub-1-meter patch cords, up to the full-spec length of 100 m.

Internet data centers introduce a nominally standard configuration for implementing server capacity into this otherwise ad hoc world. The fully populated rack, each with a pair of redundant top-of-rack (ToR) switches, is becoming a common increment of capacity. Such a fully populated rack would include the following elements:

• 42RU rack
• 40 1RU servers, each with two 10GE ports for uplinks to each of the two redundant ToR switches
• 2 ToR 1RU switches, each with 40 10GE downlinks and 8 10GE uplinks