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

I’m trying to learn CUDA and just recently started with it. As a test, I wanted to add two vectors and see if they run on the GPU. The code I wrote works, and I followed the syntax from the NVIDIA CUDA introduction. With the syntax <<<…>>>, which according to the NVIDIA guide is the kernel call, the IDE shows an error for me, indicating that it’s underlined in red and an expression is still expected. Could you help me understand why this happens? I’m sorry for this, but I’m new to it and haven’t seen anyone else online encountering the same error