Is there something even more fast than native Catboost binary?

[ibobak]

I know the question may sound naive, but we are using Java's CatBoost for inference which under the hood is using binary (.so) through JNI. The CatBoost inference (.so) is fast, but my managers want it even faster.

Recently they suggested to pay attention on this:
https://www.intel.com/content/www/us/en/developer/articles/technical/faster-xgboost-light-gbm-catboost-inference-on-cpu.html

So, I would like to ask anyone with experience in this thing:

1. is it really faster than using native CatBoost inference binary? (OK, I understand I am using through JNI, but I doubt that JNI itself takes huge %)
2. is the Intel's approach using multithreading? Once I asked here on this forum as to multithreading: I don't want this to happen in any way.

Thanks in advance.

[andrey-khropov]

1. is it really faster than using native CatBoost inference binary?

I haven't checked the performance of the aforementioned Intel library but I could try it and you should test on your particular models as well.

2. is the Intel's approach using multithreading?

Yes, but it usually can be explicitly limited (to the single thread if you like) through the library settings.