Ericsson study brings idle time into AI model energy comparisons
An Ericsson Research preprint compares eight models for predicting cellular handovers under a real-time deadline. In its simulated urban network, the energy advantage of a tree model over a neural comparator narrows when the schedule includes idle periods. The single study combines accuracy, latency and estimated processor energy, with results limited by its simulated data and tested server hardware.
Artificial Intelligence··Midday
The model must finish before the next measurement
Researchers at Ericsson Research, the telecommunications company’s research organization, compared eight models for predicting when a cellular connection should move between cells. Their single preprint examines a task with measurements arriving every 40 ms. Models must finish within that interval. The experiment checks the latency threshold containing 99 percent of cases, rather than average delay, and all models meet the limit on the tested server.[1]
Idle intervals narrow the reported energy advantage
The authors report that LightGBM, a tree-based model, has an energy advantage over CNN-1D, the strongest neural comparator, of 10.4 times when inference runs without breaks. That advantage falls to 3.3 times when inference follows a schedule with idle intervals. Accounting includes feature preparation and postprocessing as well as model execution. Feature preparation makes up about 44 percent of LightGBM’s median pipeline latency.[1]
Simulation and server measurements bound the result
The data simulates twenty-one cells and 100 devices moving for 100 seconds in a three-dimensional scene of Santa Clara in the United States. Latency tests use an isolated 36-core Xeon server processor. Energy is estimated from processor utilization sampled every 50 ms, rather than measured directly with hardware energy counters. The findings concern this simulated setting and server; they do not measure savings in a phone battery.[1]
Related columns
For more information on this topic, you can read the related columns.
