AMD developer proposes GCC patch to boost Zen 5 performance

An AMD developer, Venkataramanan Kumar, has proposed a minor patch to the GCC compiler that yields significant performance improvements for Zen 5 processors. The patch involves changing only two values within the x86-tune-costs.h header file, specifically adjusting the branch mispredict scale.

By increasing the cost assigned to branch mispredictions, the compiler is encouraged to favor branchless code variants, such as using cmov instructions, rather than risking the penalty of a mispredicted jump. In SPEC CPU 2017 floating-point benchmarks (specifically the 544.nab_r test), this change resulted in a 12% performance increase for Zen 5 and a 9% increase for Zen 4.

The reported improvements were measured using a single-copy run with -O3 -march=native -flto flags. While these results are notable, the developer noted that single-instance benchmarks represent an ideal case for branch prediction optimizations and may not reflect performance under different workloads that stress memory and cache hierarchies.