Server consolidation through virtualization is the primary strategy for optimizing data centres, aiming to maximize server efficiency and reduce costs. However, the long-term success of this strategy is compromised by three interconnected factors: virtualization overhead, the difficulty of modelling representative workloads with dependencies, and scalability limitations. This perspective article argues that scalability in consolidated environments is not a hardware problem, but rather a systems engineering problem, requiring a deep understanding of the interactions and constraints imposed by hypervisor overhead. More >

This perspective article argues that hyperparameters such as learning rate, batch size, numerical precision, and training workers are key determinants of energy scalability in CNN training. These parameters directly influence convergence dynamics, hardware utilization, and training duration, leading to substantially different energy profiles even when comparable accuracy is achieved. Moreover, hyperparameter search itself introduces a significant cumulative energy cost, often exceeding that of the final selected model. By analyzing the interaction between convergence behavior and energy consumption, this work highlights the need to treat energy as an explicit scalability metric and to integr... More >