In my testing I’ve found protobuf to be a bit faster.
Did you test against other protocols ? I've been meaning to do some tests using the c++ implementations of each used in userver.tech (specifically flatbuffers/protobuffs also unsupported sbe & captnproto) for etl processing "generally" (Again more variables are to be weighed in here for my case but would be interesting to see these things against payload sizes)
The reason we chose protobuf over capnproto was not performance but because it was already being used in other projects.
However I implemented both and didn't really do a deep comparison of performance. Basically protobuf serialisation/deserialisation was a few seconds fast. However, I didn't use capnproto as intended, in that I deserialised from the capnproto objects into an actual object which involved copying; I don't really see using the native format directly all that useful in most applications, but maybe I'm missing the point.