The article seems light on details but wouldn't writes still cause a bottleneck?
Yes, but the article talks about availability and consistency, not speed.
But it says "Replication is done in parallel with executing transactions instead of serially after commits, so it's supposed to offer better performance as well as higher consistency and availability."....
Or even more of a bottleneck if you're waiting for several writes to finish on different machines? As you say there aren't actually any details on there - might be interesting once there's another article with a bit more meat on it.