Energy-Efficiency in workflows for data-science introduces new challenges and opportunities compared with application energy efficiency. Dynamic energy-efficiency with EAR software is based on dynamic application characterization and energy policies that selects the optimal CPU and GPU frequency according to that activity.

However, the interaction with a workflow manager can be seen, from the point of view of EAR, as an expert « user », an additional monitoring and optimization layer in the software stack able to gather performance, power and energy efficiency metrics, associate with the running workflow components, and learn , and optimize the impact of energy management decisions in the overall workflow performance.

Using a synthetic workflow as an example, the presentation will highlight key challenges in moving from energy-efficient applications to energy-efficient workflows, including component-level monitoring and energy efficiency guided by external tools.

SiPearl is a European fabless semiconductor company whose aims is to design high-performance, energy-efficient microprocessors dedicated to supercomputing (HPC) and artificial intelligence (AI) workloads.

The Arm technology, used by SiPearl, has increasingly become a compelling choice for HPC due to its promise of higher efficiency; such systems are now scoring top ranks in the top500.

The presentation will guide the attendees through the Neoverse-V architecture: architectural description and specifics of SIMD extension (NEON, SVE). Some profiling and performance optimizations will be presented on Lofar workloads and in particular the fftw library.