Software-Driven Efficiency for Sustainable AI

Nikola Papadopoulou, lecturer in low carbon and sustainable computing at the University of Glasgow, argues that sustainable AI requires closing the efficiency gap between what hardware can deliver and what applications actually achieve, through adaptive system software rather than new hardware alone. She presents two systems that keep familiar programmer interfaces while making execution adaptive. The first line of work addresses communication-aware multi-GPU BLAS. Starting from single-GPU offload, performance models of data location and bidirectional link overlap drive auto-tuned tile sizing. The multi-GPU runtime extends this with topology-aware communication path optimization, tile scheduling, and device selection, achieving about 1.7x better performance and 2.5x better energy efficiency than existing auto-tuning multi-GPU BLAS libraries; a GEMM-specific variant with tile caching, batching, and bandwidth-based path selection adds a further 30-37 percent. The second system, ODIN, developed with collaborators at Chalmers and NTUA, targets interference in co-located inference workloads that use pipeline parallelism. Rather than static partitioning, ODIN treats interference as a runtime condition and rebalances layer-to-stage assignments, maintaining roughly 70 percent of interference-free throughput; a PyTorch-based successor, ReactivePipe, monitors execution and recovers 70-86 percent of initial pipeline throughput under recurring interference. Papadopoulou concludes that useful work per resource, not maximum resource usage, defines efficiency, and identifies sparse workloads, multi-model co-location, agentic AI, and energy-aware objectives as future directions. Questions address baselines, ablations, and integration with HPC schedulers.