Modern processors execute multiple instructions per cycle by predicting which code paths to follow. Branch prediction becomes critical when benchmarking performance, as processors learn patterns from repeated tests. When testing random branch conditions, processors should mispredict about half the time, but they actually learn and improve their predictions.

A simple benchmark demonstrates this learning behavior: a loop generates random values and branches based on whether they're odd or even. The processor attempts to predict these branches, and repeated runs reveal surprising learning capabilities. Three recent processors show vastly different performance levels when predicting up to 30,000 branches.

AMD's Zen 5 processor leads the pack, perfectly predicting 30,000 branches. Apple's M4 follows with 10,000, while Intel's Emerald Rapids manages only 5,000. These results highlight AMD's superior branch prediction capabilities compared to Intel's latest offerings. The source code for these benchmarks is publicly available for verification.