GC has non-zero overhead compared with linearly typed de-allocations, but that's not the main point. The title mentions predictability, and one of the big blockers to using almost any GC-ed language is the lack of predictability about runtime latency, which is very different than throughput, which I generally associate with the term bottleneck.

Additionally, predictable performance is about what the programmer can predict (whether you are talking about throughput or latency), and the argument is that linear types make it much easier to reason about the performance of a system, hence making it more predictable.