Software-Driven Efficiency for Sustainable AI

AI workloads are placing growing pressure on compute systems, and a significant amount of performance, and therefore energy efficiency, is still left on the table. Inefficient data movement, unmanaged resource contention, and software that fails to adapt to modern hardware all translate into wasted performance across the stack. This talk argues that one of the most immediate paths to sustainable AI lies in smarter software that bridges the hardware-software gap.

Drawing on my recent work across the systems stack, I will discuss how techniques such as hardware-aware autotuning and adaptive runtime management can improve both performance and energy efficiency for machine learning workloads, highlighting where efficiency is commonly lost and how systems-level optimisation can recover untapped performance in practice, supporting more sustainable and accessible AI.