I started learning cuda today and I noticed that you can either copy memory between host and device or you can use __managed__. What is the performance difference between the two and when should you use which one?

I tried both methods but I want to know the difference.