Sustainable Swarm Intelligence: Assessing Carbon-Aware Optimization in High-Performance AI Systems

Carbon-aware AI demands clear links between algorithmic choices and verified emission outcomes. This study measures and steers the carbon footprint of swarm-based optimization in HPC by coupling a job-level Emission Impact Metric with sub-minute power and grid-intensity telemetry. Across 480 runs covering 41 algorithms, we report grams CO2 per successful optimisation and an efficiency index η (objective gain per kg CO2). Results show faster swarms achieve lower integral energy: Particle Swarm emits 24.9 g CO2 per optimum versus 61.3 g for GridSearch on identical hardware; Whale and Cuckoo approach the best η frontier, while L-SHADE exhibits front-loaded power spikes. Conservative scale factor schedules and moderate populations reduce emissions without degrading fitness; idle-node suppression further cuts leakage. Agreement between CodeCarbon, MLCO2, and vendor telemetry is within 1.8%, supporting reproducibility. The framework offers auditable, runtime controls (throttle/hold/release) that embed carbon objectives without violating solution quality budgets.