Dear Developer,
I have been working with the DeepH-E3 model, specifically using the inference task[5] in the deepH-pack for sparse matrix calculations of the band structure. The system I'm working with is a transition metal dichalcogenide, with 6738 atoms and no SOC. Both elements in the system have orbital basis functions of s3p2d2. I’ve set num_bands = 40, and I’m using a machine with 128 cores and 256 threads (set with export OMP_NUM_THREADS=256 and export JULIA_NUM_THREADS=256).
However, I’m experiencing very slow performance. The computation time for one point is around 20 minutes when there are 6738 atoms without SOC, and 35 minutes when there are 9000 atoms without SOC. The process seems quite slow, especially considering I need to calculate the band structure with SOC for a system of 20,000 atoms later on.
I was wondering if there are any parameters I might have overlooked or configurations that could improve the performance. Below are my inference.ini and band.json files for reference.
Thank you for your help!
Best regards
Dear Developer,
I have been working with the DeepH-E3 model, specifically using the inference task[5] in the deepH-pack for sparse matrix calculations of the band structure. The system I'm working with is a transition metal dichalcogenide, with 6738 atoms and no SOC. Both elements in the system have orbital basis functions of s3p2d2. I’ve set num_bands = 40, and I’m using a machine with 128 cores and 256 threads (set with export OMP_NUM_THREADS=256 and export JULIA_NUM_THREADS=256).
However, I’m experiencing very slow performance. The computation time for one point is around 20 minutes when there are 6738 atoms without SOC, and 35 minutes when there are 9000 atoms without SOC. The process seems quite slow, especially considering I need to calculate the band structure with SOC for a system of 20,000 atoms later on.
I was wondering if there are any parameters I might have overlooked or configurations that could improve the performance. Below are my inference.ini and band.json files for reference.
Thank you for your help!
Best regards
