A Look At The GCC Compiler Tuning Performance Impact For Intel Ice Lake

For many different workloads, using "-march=skylake" was better off than using the actual "-march=icelake-client" or skylake-avx512 fpr that matter.

When looking at the geometric mean for the many benchmarks carried out, just using the Skylake target was about 2% faster overall than Icelake-Client or Skylake-AVX512. Only through 10 of 41 benchmarks did "-march=icelake-client" yield the fastest binary. We'll see though in a few months if anything changes for GCC 10 or under the LLVM Clang compiler when any further Icelake tuning happens or other changes. Given the similar outcome with Skylake-AVX512, presumably it's due to the AVX-512 usage leading to lower clock frequencies. If there are premium members interested in a clock monitoring comparison, let me know.