Intel Sapphire Rapids Target Added To LLVM/Clang 12.0

Intel developers engaging with upstream LLVM have been adding AMX support and other new features for next year's Xeon "Sapphire Rapids" while as of a few days ago in LLVM 12 Git is the actual enabling of -march=sapphirerapids support.

Back in July for GCC 11 the GNU compiler added -march=sapphirerapids and now similarly for LLVM 12 due out in early 2021 is also the same -march=sapphirerapids target.

Using -march=sapphirerapids for Clang on top of Ice Lake Server will enable the following extensions: amxtile, amxint8, amxbf16, avx512bf16, avx512vp2intersect, cldemote, enqcmd, movdir64b, movdiri, ptwrite, serialize, shstk, tsxldtrk, and waitpkg.

If all goes well, Sapphire Rapids will begin shipping before the end of 2021. These server CPUs will be using a 10nm Enhanced SuperFin process and are expected to come with DDR5 support, PCIe 5.0, and other new capabilities. From the compiler / software development perspective most exciting is the introduction of Advanced Matrix Extensions (AMX).

LLVM 12.0 should be out around February of next year, giving plenty of time for this compiler to be widely adopted before Sapphire Rapids CPUs hit retail channels.