|
Work item:
|
L.EE_Token
|
|
Subject/title:
|
Token-Based Evaluation Framework for Energy and Carbon Efficiency of AI Inference
|
|
Status:
|
Under study
|
|
Approval process:
|
AAP
|
|
Type of work item:
|
Recommendation
|
|
Version:
|
New
|
|
Equivalent number:
|
-
|
|
Timing:
|
2028-12 (Medium priority)
|
|
Liaison:
|
-
|
|
Supporting members:
|
China Telecommunications Corporation, Huawei, China Information Communication Technologies group
|
|
Summary:
|
This work aims to define a token-based framework for evaluating the energy efficiency and carbon efficiency of AI Inferences, including the definition of relevant terminology, fundamental measurement principles, and token-based workload accounting methods, as well as the establishment of associated energy and carbon performance indicators. The scope includes defining statistical boundaries and classification methods for input and output tokens in inference services; establishing relationships between token workload and energy consumption and defining energy efficiency indicators (such as energy per token and tokens per unit energy); proposing carbon emission calculation models based on electricity emission factors and defining related carbon efficiency indicators; and providing measurement and reporting principles applicable to different model scales, hardware architectures, and deployment configurations.
|
|
Comment:
|
-
|
|
Reference(s):
|
|
|
Historic references:
|
|
Contact(s):
|
|
| ITU-T A.5 justification(s): |
|
|
|
|
First registration in the WP:
2026-07-01 10:12:07
|
|
Last update:
2026-07-01 10:19:24
|
