AMD Zen 3 Performance With The Initial "znver3" GCC Compiler Support

For the Znver3 support at this point in GCC 11, don't expect any breakthrough changes. Normally the compiler tuning can yield up to a few percent difference in some workloads, but that is relying upon a properly tuned scheduling model and accurate cost tables for the various instructions catered to that microarchitecture family.

Across many different open-source C/C++ workloads tested, there was less than a 1% difference in Znver2 to Znver3 with the current GCC 11 code. It will be interesting to revisit when the more complete Znver3 compiler support is available but that might not land for a few months still. Some new LLVM Clang Znver3 patches were just dropped which are currently undergoing tests at Phoronix. There is also AMD's Optimizing C/C++ (AOCC) compiler that is a downstream of LLVM Clang that will be interesting to see how it performs for Zen 3 over the next release or two. At least moving off the generic x86-64 target yields around 11% uplift across all the benchmarks.

See more data over on OpenBenchmarking.org.