[kcsrk]

Better throughput, and almost no impact on latency overall. See https://arxiv.org/abs/2004.11663 section 6.2 for parallel performance evaluation.