LLVM Clang 10.0 Compiler Performance On Intel + AMD CPUs Under Linux

With the Crypto++ performance the results were basically flat except for the AVX-512 Intel CPUs. Both the Core i9 10980XE Cascadelake-X and Ice Lake systems saw better performance on Clang 10. One of the possible reasons for that improvement on the AVX-512 CPUs is Clang 10 now preferring a 256-bit vector width to limit 512-bit register usage due to the down-clocking that can often happen as a result of the lower AVX-512 clock speeds. This preferring of 256-bit matches the default behavior of GCC.

In other workloads like MrBayes we were seeing some minor improvements on Clang 10.0 for Cascade Lake and Ice Lake. In MrBayes the Zen 2 performance also pulled back slightly.

The i9-10980XE did see faster performance when built by Clang 10.

For Google's libgav1 AV1 encoder there was also slightly better performance on the i9-10980XE desktop and i7-1065G7 laptop when using Clang 10.