Currently all of the contests we have are cpu-only and never utilize parallelization (notorious exception is FHC, but even there tasks are not designed around efficient multithreading\multiprocessing). Are there any platforms already out there or in the making which have gpu-oriented small tasks and a competitive rating system?

I am aware of cuda-samples repo and competitions like kaggle, where you technically can optimize kernels for your models, but again, the repo is pretty basic and on kaggle competitions are not designed around low level optimizations like that, it's an overkill.

Wouldn't it be cool to have gpu-oriented competitions? Obviously supporting a gpu infra would be expensive for codeforces\atcoder, but maybe it can start as a "submit answers.txt" at first and then you just have to design problems properly, shouldn't cost too much to organize, but it opens up a class of new fresh problems + allows people to improve skills that are quite relevant today with all that ml hype.