The original title is: "AI Wrote a Thread-Safe Counter. The CPU Made It 5x Slower."

Short summary

An AI assistant wrote correct but slow multithreaded counter code; measurements on Apple Silicon revealed a 5x performance penalty caused by false sharing—multiple threads contending for the same CPU cache line despite updating different array elements. Padding each counter to occupy its own 128-byte cache line recovers the lost parallelism. The technique applies broadly to production code: sharded counters, cache statistics, and LRU metadata often silently suffer the same contention.

• •AI wrote correct code but 5x slower due to false sharing (multiple threads contending for same cache line)
• •Solution: pad each counter to 128 bytes so each occupies its own cache line
• •Pattern appears in production systems: sharded counters, cache statistics, LRU metadata

Generated with AI, which can make mistakes.