Mainly it will help improve latency at inference time. For preprocessing training one can always add more threads, but not at inference time.