Nginx is written in C and optimized to avoid dynamic memory allocation, whereas Go's net/http stack performs a lot of allocations, which I think are them main source of the observed performance difference (together with different approaches regarding socket handling and polling, I think). There are libraries like fasthttp that can significantly speed up Go HTTP performance, but they have their own set of tradeoffs.
So yeah if you worry about peak throughput on a CPU-bound system Caddy probably isn't the right choice though I'm not sure nginx is, either. A hardware-based load balancer, at least as a pre-stage is probably more sane in those cases. For everything else I would always trade ease of usage for a bit of performance. I would e.g. never dare to touch the nginx codebase, but writing a plugin for Caddy doesn't intimidate me at all.