コンテンツにスキップ

Benchmarks of CHGNet with GPU#

We made benchmarks of molecular dynamics (MD) calculation using CHGNet1, which is a universal graph neural network potential and available on Advance/NanoLabo, on the machine with NVIDIA H100 GPU. We executed MD calculations for sulfide-type lithium ion conductor Li10GeP2S12 and compared the results of the H100 with those of the CPU alone.

Calculation Environment#

We show the spec for the machine below.

  • CPU：AMD EPYC 9554 (64 cores)
  • GPU：NVIDIA H100
  • CUDA：12.2

HPC Systems cooperated in the preparation and use of the computing environment.

MD calculation conditions#

We generated supercell models from the structure file of Li10GeP2S12 (mp-696128) downloaded from Materials Project and executed MD calculation via MolecularDynamics class implemented in CHGNet. MD calculations were executed for 100 steps under NVT ensemble, and the time step was 0.5 fs.

Comparison of CPU and GPU#

We executed MD calculations with CPU and GPU measuring calculation time t, respectively. The results are shown in the table and the figure below. Also, we show ratio of CPU to GPU calculation time in the table. Besides, we set number of threads as 1 in any case of the conditions.

We can see that GPU is about 7 to 8 times faster than CPU from the table. However, for 50 atom system, calculation speed of GPU relative to CPU is only about 3.54 times faster. This suggests that GPU do not perform fully when the system size is small.

tCPU × 1 (sec) tGPU × 1 (sec) tCPU × 1/tGPU × 1
50 atoms 11.52 3.25 3.54
400 atoms 70.80 8.39 8.44
1,350 atoms 244.76 31.15 7.86
3,200 atoms 615.96 85.76 7.18
6,250 atoms - 172.25 -
10,800 atoms - 293.91 -
12,600 atoms - 349.90 -
13,500 atoms - 367.34 -

We can see that calculation time increases almost linearly with the number of atoms in both CPU and GPU cases from the figure. This indicate that GPU may be about 7 to 8 times faster than CPU for systems including more than 3,200 atoms. We note that the calculation for the 14,700 atom system stopped due to out of memory in the calculation environment.

The results of benchmarking show that the MD calculation using CHGNet with NVIDIA H100 GPU can be performed about 7 to 8 times faster than with CPU alone.

On Advance/NanoLabo, an integrated GUI for nanomaterials, CHGNet is available with GPU, so it will enable you to execute MD calculation using CHGNet more shortly.

関連ページ#

  1. Deng, B., Zhong, P., Jun, K. et al. CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nat Mach Intell 5, 1031–1041 (2023). ↩