I made a kernel 2.2x faster. It made my training loop 3x slower

In three linesA developer optimized a kernel by 2.2x but this made their training loop 3x slower. The post illustrates the common optimization paradox: improving an isolated component can degrade overall performance due to hidden bottlenecks, memory pressure, or latency shifts.

Summary generated by Claude — human-verified