Exploring Computer Architecture Designs Tailored to Next Generation Computational Drug Discovery
Wednesday, June 24, 2026 3:45 PM to 5:15 PM · 1 hr. 30 min. (Europe/Berlin)
Foyer D-G - 2nd Floor
Project Poster
Bioinformatics and Life SciencesNovel AlgorithmsPerformance and Resource ModelingPerformance MeasurementPerformance Tools and Simulators
Information
Poster is on display.
The IMPACT project closely aligns with the ICON Health Mission and addresses challenges of significant societal relevance. Its primary objective is to improve the efficiency of early-stage drug discovery for small-molecule therapeutics by advancing the use of machine learning techniques, quantifying the future computational requirements of increasingly accurate but computationally expensive simulation methods, and exploring how imec can automate and scale the generation of high-quality data required to train data-driven machine learning models. By reducing cost and development time and enabling the exploration of a larger chemical space, the project has the potential to substantially accelerate therapeutic discovery while lowering experimental burden.
Molecular dynamics (MD) simulations model the time evolution of atomic systems governed by interatomic forces, and their predictive fidelity depends critically on the accuracy of the underlying force models. Classical force fields (CFFs) rely on fixed functional forms parameterized from experimental or theoretical data, offering computational efficiency and broad applicability but limited accuracy in chemically complex or reactive environments. In contrast, machine learning force fields (MLFFs) achieve near quantum-chemical accuracy at molecular-mechanics cost by learning interatomic interactions directly from high-level electronic-structure data.
Although MLFFs substantially improve accuracy compared to classical approaches, they introduce significant computational overheads, particularly in descriptor construction and neural network inference. These workloads pose challenges for parallel hardware due to irregular memory access patterns, limited data reuse, and inefficient kernel execution. One of the core goals of the IMPACT project is to characterize the hardware performance of MLFF-based MD simulations using poly-alanine chains as a novel benchmark system with controllable input size. This benchmark enables systematic evaluation of GPU scalability and identification of key computational bottlenecks encountered when scaling MLFF simulations.
MD simulations rely on a set of computationally intensive kernels, including non-bonded force evaluations, long-range electrostatics, and constraint solvers. Owing to their high floating-point operation density, irregular memory access behaviour, and dependence on massive parallelism, MD workloads place substantial pressure on both compute and memory subsystems, often becoming limited by memory bandwidth and cache inefficiencies.
As hardware architectures evolve toward increased heterogeneity and specialization, it is essential to evaluate how MD kernels will perform on future systems prior to hardware availability. A secondary objective of the IMPACT project is therefore to assess selected MD kernels using architectural simulation platforms to analyse their performance on simulated next-generation hardware. Through detailed exploration of architectural parameters such as memory hierarchy, SIMD width, and interconnect design, combined with targeted tuning of system and kernel configurations, this work aims to support hardware–software co-design strategies for accelerating MD simulations.
Hierarchical roofline-based analytical performance models are employed as an initial architectural pathfinding step to identify bottlenecks and to assess the impact of future system design choices on next-generation MD engines. Developing analytical performance models that capture the interaction between MD algorithms and compute-architectures represents an additional key objective within the scope of the IMPACT project.
The IMPACT project closely aligns with the ICON Health Mission and addresses challenges of significant societal relevance. Its primary objective is to improve the efficiency of early-stage drug discovery for small-molecule therapeutics by advancing the use of machine learning techniques, quantifying the future computational requirements of increasingly accurate but computationally expensive simulation methods, and exploring how imec can automate and scale the generation of high-quality data required to train data-driven machine learning models. By reducing cost and development time and enabling the exploration of a larger chemical space, the project has the potential to substantially accelerate therapeutic discovery while lowering experimental burden.
Molecular dynamics (MD) simulations model the time evolution of atomic systems governed by interatomic forces, and their predictive fidelity depends critically on the accuracy of the underlying force models. Classical force fields (CFFs) rely on fixed functional forms parameterized from experimental or theoretical data, offering computational efficiency and broad applicability but limited accuracy in chemically complex or reactive environments. In contrast, machine learning force fields (MLFFs) achieve near quantum-chemical accuracy at molecular-mechanics cost by learning interatomic interactions directly from high-level electronic-structure data.
Although MLFFs substantially improve accuracy compared to classical approaches, they introduce significant computational overheads, particularly in descriptor construction and neural network inference. These workloads pose challenges for parallel hardware due to irregular memory access patterns, limited data reuse, and inefficient kernel execution. One of the core goals of the IMPACT project is to characterize the hardware performance of MLFF-based MD simulations using poly-alanine chains as a novel benchmark system with controllable input size. This benchmark enables systematic evaluation of GPU scalability and identification of key computational bottlenecks encountered when scaling MLFF simulations.
MD simulations rely on a set of computationally intensive kernels, including non-bonded force evaluations, long-range electrostatics, and constraint solvers. Owing to their high floating-point operation density, irregular memory access behaviour, and dependence on massive parallelism, MD workloads place substantial pressure on both compute and memory subsystems, often becoming limited by memory bandwidth and cache inefficiencies.
As hardware architectures evolve toward increased heterogeneity and specialization, it is essential to evaluate how MD kernels will perform on future systems prior to hardware availability. A secondary objective of the IMPACT project is therefore to assess selected MD kernels using architectural simulation platforms to analyse their performance on simulated next-generation hardware. Through detailed exploration of architectural parameters such as memory hierarchy, SIMD width, and interconnect design, combined with targeted tuning of system and kernel configurations, this work aims to support hardware–software co-design strategies for accelerating MD simulations.
Hierarchical roofline-based analytical performance models are employed as an initial architectural pathfinding step to identify bottlenecks and to assess the impact of future system design choices on next-generation MD engines. Developing analytical performance models that capture the interaction between MD algorithms and compute-architectures represents an additional key objective within the scope of the IMPACT project.
Format
on-demandon-site