GPU support

[BradKML]

If this model is fast, then would including the usage of GPU makes the speed even faster (for larger datasets)? Think LightGBM does have GPU support https://lightgbm.readthedocs.io/en/latest/GPU-Tutorial.html
The only problem I can see with this arrangement, would be that it is written in Rust, so the low-level stuff and the GPU stuff might mingle more oddly

[deadsoul44]

GPU programming requires specific libraries. I don't have that expertise, but I'm happy to help with algorithm details if someone wants to tackle the GPU implementation.

[serkor1]

@deadsoul44 I am sure you have seen this repository: https://github.com/NVlabs/cuda-oxide, but I just wanted to link it anyways.

The documentation is proper, and it looks like an implementation that is worth pursuing without having to refactor alot of the backend. We could give it a year or two until it gets released as stable.

[deadsoul44]

Thanks for the heads-up! I hadn't seen this repository, but it looks like a solid candidate for prototyping. I’m not overly concerned about the library being in an experimental phase. The bigger question is the current demand for GPU acceleration in PerpetualBooster. If the need is there, this seems like a very viable path to explore.

[serkor1]

I agree that it could be too preemptive to implement it now. Either way, my plan of attack would be to start an entirely new repository testing it out on a simple algorithm to get familiar with cuda-programming - especially considering that we are starting from zero.

In essence we could make a quick repo with a simple linear regression model that has the same underlying logic as PerpetualBooster (pointers, parallel computing, etc) and see how well it works with cuda, that way we could probably make a better plan of attack when the time comes. I have a feeling that GPU support is a huge endeavor.

[deadsoul44]

Good idea to try it with a more simple algorithm. It will enable us to better evaluate how to tackle it for PerpetualBooster.

[serkor1]

I have played around with it - it looks easy at a first glance, but then I tried to scale it to a more elaborate routine and even with AI it becomes really complicated.

I thought it was Rust with some extra declarations, but its more like rewritten CUDA in Rust.

[deadsoul44]

Then, probably we will have to re-write the most of the algorithm if we use this library. Directly using CUDA might be more reasonable.