Speaker
Description
Barbora NG is the latest CPU-based supercomputer operated by IT4Innovations, consisting of 144 nodes, each equipped with two Intel Xeon 6952P processors. Benchmarking its sustainable performance is therefore an important step in assessing its energy efficiency and identifying potential performance bottlenecks of this new HPC platform.
In this work, we utilize a Roofline benchmark with a set of predefined arithmetic intensities covering both memory-bound and compute-bound regimes. The benchmark is instrumented with the MERIC Toolkit, enabling monitoring of core and uncore frequencies together with power draw. This provides a more detailed analysis of performance behavior and potential bottlenecks beyond the conventional metrics of computational performance in FLOP/s and memory throughput in GB/s. The benchmark results and MERIC measurements are further used to quantify CPU-to-CPU variability in both performance and power draw across the Barbora NG HPC platform.