Why is the deepmd-kit so fast?

[kudoryutaro]

Hello.

I'm making DeepPot-SE AtomTypeEmbedding version using pytorch.

So my question is, why is deepmd-kit so fast?

I implemented the forward and backward of the R_i -> R^Tilda_i part by hand using pytorch.

However, it takes 0.1 sec to infer the force and potential energy with 5000 atoms at a time.

On the other hand, deepMD-kit can calculate about 40000 atoms in 0.1 sec. Why is it so fast?

[wanghan-iapcm]

Hello,

I cannot tell you the different unless we have detailed benchmarks for both implementations. One possibility is the difference of deep learning platforms. DeePMD-kit uses tensorflow v1 which uses static graph, while pytorch uses dynamic graph.