Optimizing AI data center power efficiencies

AI-optimized data center networks offer significantly higher bandwidth —100 to 1000 times more than a Traditional workloads, but this increase comes with substantial power demands.

To address these demands, the focus is on reducing overall system power consumption, particularly by decreasing optics power consumption, which can account for approximately 50% of the total power used in AI networking.

Cisco's Silicon One includes a world-class SerDes (serializer/deserializer) component that is used to convert parallel data into serial data and vice versa. Its unique capabilities allow for the use of lower power linear pluggable optics (LPO) or passive direct attach copper cables (DAC), thereby reducing power consumption for a sustainable AI/ML cluster.

Traditional retimed optics

In a traditional retimed 800 Gbps optical module used in a 51.2 T switch there are two main transmission components:

  • The optical front end is responsible for converting electrical signals to and from optical signals, enabling efficient data transmission over fiber.
  • The DSP-based retimer enhances signal quality by retiming the electrical signals as they travel to and from the fiber.

Together, these components contribute to significant system power consumption, with 64x optics accounting for approximately 1000W of system power in such a high-capacity switch.

Linear pluggable optics (LPO)

In Cisco Silicon One, the DSP-based retimer functionality is integrated directly into the G200 switch ASIC, while the optical device is focused solely on the optical front end.

Converting all 64 optics saves approximately 500W of system power for a 51.2T switch, significantly lowering power consumption.

Direct attached cable (DAC)

A passive copper cable, also known as a Direct Attach Copper (DAC) cable, enables 0W connectivity within a rack, even at speeds of 800G.

By converting 64x optics to DAC, a 51.2T switch can save approximately 1000W of system power.

This shift reduces the overall power consumption significantly, making it a highly effective solution for short-distance connections in AI workloadss.