VASP¶

VASP (Vienna Ab initio Simulation Package) performs ab initio quantum-mechanical molecular dynamics using pseudopotentials and a plane-wave basis set.

Licensing¶

VASP requires a valid license. Send your license details to support-hpc@dipc.org so we can verify it and grant you access to the binaries.

Loading the module¶

module load VASP/6.4.3-intel-2024a

Each module provides three binaries. Choose the one that matches your calculation:

• vasp_std -- General purpose. Default choice for most calculations.
• vasp_gam -- Gamma-point only. Uses real arithmetic instead of complex, roughly 2x faster and half the memory. Use it when your KPOINTS file has only the Gamma point.
• vasp_ncl -- Non-collinear magnetism and spin-orbit coupling. Required for SOC calculations; slower than vasp_std.

To see all installed versions across clusters, check the software catalog or run module spider VASP.

Pseudopotentials¶

VASP pseudopotential files (POTCARs) are distributed as part of the VASP package and access is controlled by your license. If your license covers VASP 6, you can access the VASP 6 pseudopotentials; if it covers VASP 5, you can access the VASP 5 set.

Batch job examples¶

#!/bin/bash
#SBATCH --qos=regular
#SBATCH --job-name=vasp
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=48
#SBATCH --cpus-per-task=1
#SBATCH --mem=200G
#SBATCH --time=24:00:00
#SBATCH --output=%x.%j.out
#SBATCH --error=%x.%j.err

module load VASP/6.4.3-intel-2024a

srun vasp_std

#!/bin/bash
#SBATCH --qos=regular
#SBATCH --job-name=vasp_gpu
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=1
#SBATCH --mem=40G
#SBATCH --gpus=1
#SBATCH --time=24:00:00
#SBATCH --output=%x.%j.out
#SBATCH --error=%x.%j.err

module load VASP/6.4.1-NVHPC-23.1-CUDA-12.1.1

export OMP_NUM_THREADS=1
srun vasp_std

#!/bin/bash
#SBATCH --qos=regular
#SBATCH --job-name=vasp
#SBATCH --nodes=4
#SBATCH --ntasks-per-node=48
#SBATCH --cpus-per-task=1
#SBATCH --mem=200G
#SBATCH --time=24:00:00
#SBATCH --output=%x.%j.out
#SBATCH --error=%x.%j.err

module load VASP/6.4.1-intel-2022a

srun vasp_std

GPU builds¶

VASP 6.4+ supports GPU acceleration via OpenACC. On Hyperion, for example, a GPU-enabled build is available:

module load VASP/6.4.1-NVHPC-23.1-CUDA-12.1.1

This build is compiled with NVHPC 23.1 and CUDA 12.1, linked against MKL for CPU math and cuFFT/cuSOLVER for GPU-accelerated FFTs and linear algebra. All three binaries (vasp_std, vasp_gam, vasp_ncl) are GPU-enabled. Check module spider VASP on your cluster for other GPU-enabled modules (look for NVHPC or CUDA in the name).

Key differences when running on GPU:

• Use one MPI rank per GPU (e.g., --ntasks-per-node=4 with --gres=gpu:4)
• Set NCORE = 1 in your INCAR (mandatory for GPU; see VASP wiki: OpenACC GPU port)
• Set OMP_NUM_THREADS=1

Performance tuning (INCAR)¶

The following INCAR parameters have the biggest impact on performance. Incorrect settings are the most common reason for slow VASP runs.

Example INCAR block for a 2-node CPU job on Hyperion (96 MPI ranks):

NCORE  = 8
KPAR   = 2
LREAL  = Auto
NSIM   = 4

Scaling guidelines¶

How many cores your calculation needs depends primarily on the system size and the number of k-points. These are starting points; always verify by comparing the LOOP+ timing in the OUTCAR between different configurations.

How to run a scaling test¶

Run the same calculation on 1, 2, and 4 nodes with a small number of electronic steps (NELM = 5) and compare the elapsed time per step (LOOP+ in OUTCAR):

grep "LOOP+" OUTCAR

If doubling the nodes does not reduce the time by at least 30-40%, you are past the efficient scaling range. Use fewer nodes and spend the saved core-hours on more calculations instead.

For a detailed discussion of VASP parallelization strategy, see the official optimization guide.

See also¶