So due to how Go handles memory, it needs to do something called "trampoline", over to the C stack. This causes Go->C calls to have a substantial overhead 70-200ns/op. Most other languages that value is closer to 1ns. This means that every call to a GPU suffers from this latency.

This isn't too much of a problem if your doing all supervised learning with batch operations because the speedup of a GPU over a bigger operation outweighs the FFI latency.

However, it's a problem that doesn't appear to have a solution due to Go's memory management choices, and will hamper it ever being used for accelerated computing problems. This is one of the reasons Rust moved to using ownership rules.

You can read a bit more at https://dave.cheney.net/2016/01/18/cgo-is-not-go