Is it heavy on the GPU to write data to the RWStructuredBuffer?

Question

C++ / C#

In my practice, it loads very, very much even (the structure being changed contains about 16 floats). Simply equating each element to one reduces the fps from 1500 to 300. But then how do you work with RWStructuredBuffer if writing to it is so expensive? At the same time, in this article https://habr.com/en/post/461017the author says: "Also of concern is the massive I/O of buffers. This is indeed a difficulty, but not as serious as one might expect: data access is highly predictable, especially when writing to buffers, so latency is not a problem. Actually , for this type of data processing, GPUs were primarily developed.". Please explain how it works and how it can be optimized, well, or throw off a link to the source, where you could learn more about it.