HiDALGO2 team expands our understanding of how modern hardware and software innovations can optimise large-scale environmental simulations and AI-driven workflows. D3.6 deliverable: Innovative HPC Technologies and Benchmarking, currently under review by the EC, builds on our previous evaluations, as described in our previous reports D3.4 and D3.5 and focuses on state-of-the-art architectures from the world’s leading processor manufacturers, assessing their raw performance, communication efficiency, and energy footprint.
Driving Environmental Computing with Next-Gen CPUs
To ensure our findings translate directly to real-world impact, Deliverable 3.6 evaluates the latest hardware using two highly demanding use cases critical to environmental computing: the Wildfires pilot (utilising the coupled fire-atmosphere WRF-SFIRE model) and Sparse Large Language Model (LLM) inference.
Our benchmarking testbeds represent the absolute vanguard of high-performance processing, featuring:
– AMD EPYC 9004 & 9005 Series: Evaluated across multiple generations—including standard Genoa, cache-expanded Genoa-X (utilizing 3D V-Cache), next-generation Turin (Zen 5 architecture), and high-density Turin Dense (Zen 5c architecture) up to an incredible 192 cores per socket.
– Intel Xeon 6 (Granite Rapids): Highlighting Intel’s flagship 6980P Performance-core (P-core) model operating with 128 cores per socket.
Key Takeaways from Our Benchmarking & Profiling
1. The Sweet Spot for Hybrid MPI+OpenMP Parallelism
One of the core investigations of this report was identifying how to balance Message Passing Interface (MPI) processes with OpenMP threads to extract maximum efficiency.
– The Findings: Across all evaluated AMD platforms and scale configurations, the optimal runtime and energy efficiency minimum was consistently found at 4 OpenMP threads per MPI rank.
– Why it Matters: Adopting an optimal hybrid balance reduces the total number of MPI processes, slashing point-to-point communication and halo data exchange volumes by roughly 60% compared to pure MPI configurations. This optimization effectively mitigates synchronization bottlenecks as systems scale.
2. Performance vs. Energy Efficiency: Turin Leads the Way
Using the AMD µProf performance profiling toolkit, our teams thoroughly explored how microarchitectural changes directly translate to energy savings:
– Turin (Zen 5) Superiority: The dual 128-core AMD EPYC Turin platform emerged as the standout architecture, delivering the shortest execution times and achieving excellent computational throughput (approaching 500 GFLOPS).
– The Energy Balance: While newer generations like Turin often demand a higher average power envelope under full load, their significantly shorter runtimes mean they consume 15–25% less total energy (kWh) to complete the same workload compared to previous-generation setups.
– Diminishing Returns at Extreme Densities: Interestingly, pushing core density further with the 192-core Turin Dense (Zen 5c) did not yield better performance for WRF-SFIRE. This indicates that memory-bound environmental stencils benefit far more from robust per-core cache allocations and higher frequencies than simply maximizing raw core count.
3. Breaking Bottlenecks in Sparse LLM Inference
As AI models increasingly merge with environmental science workflows—such as predicting fire spread variations or simulating urban pollutant dispersion—understanding CPU-based LLM inference efficiency is important.
– By substituting dense mathematical matrices with Sparse Matrix-Matrix Multiplication (SpMM) and Sampled Dense-Dense Matrix Multiplication (SDDMM), our researchers minimized memory overhead.
– While a theoretical schedule concurrently executing three SpMM kernels across the transformer pipeline indicates an average 20% reduction in total execution time, our empirical testing exposed library-level limitations. Current vendor library constraints, such as static thread affinities and scheduling overheads, show that custom, pipeline-aware kernels are required to fully unlock concurrent execution.
This deliverable is the outcome of a joint effort of key researchers Lukasz Szustak and Marcin Lawenda from PSNC (Poznan Supercomputing and Networking Center) and Ioanna Tasou, Petros Anastasiadis, and Konstantinos Nikas from ICCS (Institute of Communication and Computer Systems) who made this comprehensive study possible. It was reviewed by Luis Torres, Leydi Laura Salazar, Angela Rivera and Jaime Ribalaygua from MTG (Meteogrid), Abdullatif Eymash from USTUTT (University of Stuttgart) and Shubham Kavane (Quality Manager) FAU (Friedrich-Alexander-Universität Erlangen-Nürnberg).
Stay tuned as we continue to work on innovative hardware technologies to deliver faster, cleaner, and more precise environmental simulations!
👉 Read the Full Deliverable on Zenodo
