GPU Acceleration for Faster Evolutionary Spatial Cyclic Game Systems

TL;DR

This paper develops GPU-accelerated simulation frameworks for Evolutionary Spatial Cyclic Games, significantly improving computational speed and enabling larger system sizes, thus advancing ecological and evolutionary modeling capabilities.

Contribution

It introduces high-performance GPU implementations of ESCGs using CUDA and Metal, achieving up to 28x speedup and enabling simulations of larger systems than previously possible.

Findings

GPU acceleration yields up to 28x speedup in ESCG simulations

Larger system sizes up to 3200x3200 are now feasible

GPU frameworks facilitate replication and extension of existing studies

Abstract

This dissertation presents the design, implementation and evaluation of GPU-accelerated simulation frameworks for Evolutionary Spatial Cyclic Games (ESCGs), a class of agent-based models used to study ecological and evolutionary dynamics. Traditional single-threaded ESCG simulations are computationally expensive and scale poorly. To address this, high-performance implementations were developed using Apple's Metal and Nvidia's CUDA, with a validated single-threaded C++ version serving as a baseline comparison point. Benchmarking results show that GPU acceleration delivers significant speedups, with the CUDA maxStep implementation achieving up to a 28x improvement. Larger system sizes, up to 3200x3200, became tractable, while Metal faced scalability limits. The GPU frameworks also enabled replication and critical extension of recent ESCG studies, revealing sensitivities to system size and runtime not fully explored in prior work. Overall, this project provides a configurable ESCG simulation platform that advances the computational toolkit for this field of research. This dissertation forms the basis for a paper accepted for publication and presentation at the European Modelling and Simulation Symposium.