Depending on your needs, I have generally found that between Fastmath and Neanderthal, I have no real worries about performance sensitive numeric code. Obviously if you start caring about CPU caches and specific instructions then you might struggle (I am writing a chess engine in Clojure for no good reason, and obviously it's not going to hit the level of throughput that engines written in C++ reach). There are annoying niggles, like Clojure not supporting all Java's primitive types, and obviously the wider support ecosystem for data science work is somewhat weak, but I've been impressed.