Workshop: AI on HPC: Performance Engineering, Challenges and Opportunities
Authors: Philipp Bader (Technical University of Munich); Jophin John (Leibniz Supercomputing Centre); and Eishi Arima, Michael Gerndt, and Martin Schulz (Technical University of Munich)
Abstract: Datacenter electricity consumption is projected to grow rapidly, driven primarily by AI. Meeting this demand sustainably requires understanding how training energy scales with model size, and whether this holds across increasingly diverse hardware, particularly for wafer-scale dataflow accelerators like the Cerebras WSE series, which integrate compute and memory on a single wafer, unlike die-scale GPUs. We compare a Cerebras WSE-2 and an NVIDIA GH200 as representative wafer- and die-scale accelerators, using a matched sweep of seven CNN and Vision Transformer architectures with batch sizes from 2 to 512 to cross-validate energy models. We find distinct energy-scaling behavior between the two: the WSE-2's dataflow execution makes scaling more predictable, while its high baseline power and lack of workload-controlled power management increase sensitivity to model size and batch size.
Back to AI on HPC: Performance Engineering, Challenges and Opportunities Archive Listing
Back to Full Workshop Archive Listing