Comparison/Improvements of cuEquivariance over e3nn

[duydl]

Hi,
Thank you for sharing this project! Could you provide a brief comparison or outline any improvements cuEquivariance offers over the established e3nn library? Specifically, I’m curious about performance, flexibility, and API design.

Thanks for your time and work!

[hiranumn]

I am curious to hear about this as well.

[becca]

There a few notable differences between e3nn and cuEquivariance, namely around performance, flexibility and API design.

1. MACE with cuEquivariance have notable speedups, some of which we published in https://developer.nvidia.com/blog/accelerate-drug-and-material-discovery-with-new-math-library-nvidia-cuequivariance/. Additionally, MACE is now integrated with cuEquivariance (Add cuequivariance support ACEsuit/mace#704) 🎉
2. e3nn is bound to a specific choice of SO3 irreps basis and data layout, in cuEquivariance you can specify your own irreps basis and tensor product by creating a segmented tensor product.
3. We provide the new features mentioned above with an API that is very similar to that of e3nn.

Let us know if you have any other questions or feedback.