Clear Linux Discovers Another AVX2/AVX512 Fix/Optimization To Yield Better Performance

For those running a system with AVX-512 support, Clear Linux builds as of this week should be yielding even better performance on top of their existing AVX2 and AVX-512 optimizations.

The Intel developers working on Clear Linux uncovered an issue how the new GCC 9 compiler has been building the important libm math library poorly in AVX2/AVX-512 mode. This poor code compilation yielded slowdowns in various math functions since the switch to the GCC 9 compiler.

As of the Clear Linux builds since Wednesday, the issue should be addressed. Meanwhile the Intel developers are working on a proper upstream fix as well to GCC/glibc, benefiting other Linux distributions as well when moving to newer versions of these important toolchain components.

I looked at some tests briefly using an AVX512-enabled Xeon Silver 4108 and indeed in many workloads there are nice improvements to be found.


By far the biggest winner of this fix/optimization was Smallpt:

I'll be trying out more tests of the latest Clear Linux on other systems shortly.