Artificial intelligence can deliver public value without locking institutions into ever-rising energy demand. This research examines architectures that treat efficiency as a first-order design constraint.
Our benchmark combines structured sparsity, mixed-precision quantization, and workload-aware scheduling. Across representative edge workloads, the approach lowers measured energy use while preserving task performance.
The findings support a broader conclusion: transparent measurement and reusable infrastructure must become standard parts of responsible model development.
Research implications
This work is published to support replication, constructive critique, and responsible application. Research partners may contact GSAI for methods, collaboration, and accessible formats.