Workshops Archive

Energy and Cost-Aware Control at Scale: Multi-Dimensional Frequency Steering on AMD MI300A


Workshop: EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices

Authors: Christian Simmendinger, Marcel Marquardt, and Elias Klett (HPE) and Ralf Schneider (HLRS)

Abstract: Energy and Power have become a first-order concern for large HPC and AI installations. The optimization targets are no longer limited to production peak throughput, but include stable output under electrical limits, carbon constraints, and infrastructure cost pressure. This paper presents an extension to an operational control approach (PowerSched) that mitigates some of these concerns. The work reports on a PowerSched extension towards multi-dimensional frequency steering of GPU architectures that targets CPU boost frequencies, GPU clocks and memory/interposer-related controls, The control plane supports policy shaping through a configurable objective term, so operation can be shifted at runtime along an energy--performance--cost trade-off axis. Challenges and solutions for multi-dimensional steering algorithms are discussed, including mutual constraints, discrete-valued tunables, and noisy telemetry. First results for AMD mi300a architecture are presented. The results indicate performance gains up to 10% relative to the uncapped system and energy-efficiency gains up to 30%.


Back to EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices Archive Listing

Back to Full Workshop Archive Listing