This repository contains a few benchmarks for YASTN.

We follow a benchmark framework adopted by the authors of TeNPy and ITensor. It considers a spin-1 Heisenberg chain of 100 sites, performing a number of 2-site DMRG sweeps with increasing MPS bond dimensions.

See, TeNPy benchmarks and TeNPy and ITensor Comparisons.

After installing YASTN and cloning this repository, our benchmark script can be run as

python bench_dmrg.py

A few control options are available from the command line

python bench_dmrg.py --help

We run this benchmark focusing on U(1)-symmetric tensors employing a workstation with Intel i9-10920 CPU and Nvidia RTX-3090 GPU.

We first run the test on a single CPU core to compare the execution times of the three libraries, obtaining comparable performance for the corresponding simulation setups. Those employ package versions are: YASTN v1.2, TeNPy v1.0.3,  ITensors v0.6.16.

Next, we run YASTN using NumPy and PyTorch backends across one and multiple cores and utilizing GPU. PyTorch tensors show systematic overhead visible for small bond dimensions, which becomes marginal for larger bond dimensions. For CUDA (GPU) computation, we delegate the SVDs to the CPU due to the poor performance of SVD decomposition on the GPU. Still, in this example, the timings of large bond-dimension sweeps employing GPU and multiple-core CPUs are dominated by the SVD.

We thank TeNPy developer Johannes Hauschild for the discussions.

We collect core elements of CTMRG update for tensor sizes motivated by the applications described in YASTN release article. See .\bench_contractions.py for details of the test framework, with symmetric tensor structures gathered in the folder .\input_shapes\.