"Tracing garbage collection has better throughput than reference counting, but reference counting has lower and more predictable latency."

Exactly this. I would actually back off a bit and only claim that reference counting has more predictable latency, but the point is to sacrifice throughput in order to reduce worst case latency.