Workshop: Sustainable Supercomputing
Authors: Tim Thüring, Malte Horst, Alexander Strack, Alexander Van Craen, and Dirk Pflüger (University of Stuttgart)
Abstract: Recently, the landscape of heterogeneous HPC systems has changed drastically once more, from classic GPU acceleration to tightly coupled systems such as the NVIDIA Grace Hopper superchip and the AMD MI300A APU. Since energy consumption is a critical factor for the sustainable operation of HPC facilities, we analyze a heterogeneous SYCL application capable of CPU-GPU co-execution to investigate how power draw, energy consumption, and performance evolve in this ongoing transition. Comparing heterogeneous and GPU-only scenarios, we find that the trade-off between maximum performance and minimum energy consumption observed in a conventional system persists even in modern superchips. Furthermore, we observed significant energy savings in GPU-only settings on NVIDIA's GH200 platform compared to a classic GPU-accelerated system, driven by an over 140W lower CPU idle power draw. In a GPU-only scenario, AMD's MI300A APU achieves almost 19 percent lower runtimes than the NVIDIA GH200, while consuming only 1.03 times the energy.
Back to Sustainable Supercomputing Archive Listing
Back to Full Workshop Archive Listing