Yes, I confirm this issue as well. I get accurate and better results when running the same exact code in Nvidia on Windows. Seems like Cuda isn't yet optimized for Apple silicon, and most likely it will never be.