started · updated
AI workloads drive shift toward energy per inference metric in data centers
The data center industry is transitioning from using Power Usage Effectiveness (PUE) as its primary efficiency metric toward a more granular measure: energy per inference. This shift is driven by the rapid scaling of artificial intelligence (AI) workloads, which require more specific insights into how efficiently AI tasks are executed rather than just facility-wide power consumption.
As data rates increase to 800G and 1.6T, the energy required to move data via optical interconnects is becoming a critical factor in total power consumption. Technologies such as Linear Pluggable Optics (LPO) are being utilized to improve efficiency. By removing the power-intensive Digital Signal Processor (DSP) and shifting signal conditioning to the host switch silicon, LPO can reduce the power draw per 800G link from approximately 13W-16W to between 7W and 9W.
Separately, traditional PUE remains a standard for measuring total site energy versus IT equipment energy. A PUE of 1.20 is generally considered excellent, while scores above 1.70 are viewed as needing improvement.
Entities
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 2 SOURCES] In standard full re-timed optics, the Digital Signal Processor (DSP) accounts for approximately 40% of total optic power consumption. bbcmag.com · totaltele.com
- [● 2 SOURCES] Linear Pluggable Optics (LPO) can reduce the power draw per 800G link from 13W-16W down to 7W-9W. bbcmag.com · totaltele.com
- [● 2 SOURCES] The data center industry is shifting its focus from Power Usage Effectiveness (PUE) to the metric of energy per inference due to scaling AI workloads. bbcmag.com · totaltele.com
- [○ 1 SOURCE] A Power Usage Effectiveness (PUE) score below 1.20 is considered excellent. alice-ecologie.fr