TPU megakernel

A deep-learning inference optimization technique that merges multiple discrete operator calls into a single large kernel. By reducing kernel launch overhead, memory round-trips, and scheduling latency, it improves hardware utilization and throughput.