Supervisors: Mehran Salmani, Boris Koldehofe
Problem Description:
Deep learning (DL) models have predictable latency. Knowing the latency of the DL model when running on specific hardwares helps optimizing the resource allocation and energy usage of DL models to meet the latency requirement as a service level objective.
This thesis is about performance and power consumption prediction of variable DL models on variable (unseen) GPUs considering colocation.
Requirements:
Knowlege of operating systems, deep learning model structure, and GPUs