Overall GPU performance and scalability measurements

[etiennemlb]

Summary

Hi,

I am currently looking for some performance metrics regarding DeepMD, specifically:

1. Its scalability, particularly in terms of network communications. If you have any data on scalability efficiency measurements (strong/weak scaling), that would be incredibly helpful.
2. Its ability to utilize hardware effectively, including high FLOPS and power usage. Any roofline models or power usage traces over time would be greatly appreciated.

I understand that these numbers can vary depending on the specific use case, but I am just trying to get a general sense of its capabilities.

DeePMD-kit Version

take the latest

Backend and its version

assume the latest

Python Version, CUDA Version, GCC Version, LAMMPS Version, etc

No response

Details

We are considering DeepMD for use in HPC environments and are keen to ensure that we can make the most out of our hardware resources. The computational demands of the project I have in mind are quite significant, so it's important for us to gauge how well DeepMD might perform in such a setting.

Of course, we plan to conduct our own benchmarks, but as a preliminary step, any insights you could provide would be very valuable.

Thank you very much for your time and assistance.

Cheers.

[njzjz]

There was a scaling benchmark published in 2020: https://doi.org/10.1016/j.cpc.2020.107624 (Fig 8-9)

[wanghan-iapcm]

The scaling benchmark for the DeepPot-SE model was established in 2020, as @njzjz noted.

Currently, we lack scalability and hardware efficiency benchmarks for recently developed graph models, such as DPA-2/3. We warmly invite and appreciate any contributions you may have in reporting the scalability of these models.