GPU Scaling for Materials

Large-scale materials simulation is only useful if it runs fast and stays numerically trustworthy. This project benchmarks two workloads on NVIDIA Blackwell B200 GPUs: ML-force-field (ALIGNN-FF) inference on copper supercells from a few atoms up to ~780k atoms, and VASP DFT self-consistent-field scaling on silicon supercells. Along the way it characterizes a float32 energy-drift “cliff” near ~470k atoms that bounds the safe single-precision operating window, and measures strong- and size-scaling behavior to inform how big a simulation can responsibly go.