NVIDIA has upstreamed a patch to the GCC 15 compiler for adding the "-mcpu=grace" option to make it easier to target NVIDIA Grace AArch64 CPU cores from this open-source compiler.

This patch in GCC 15 Git as of today now allows for enabling all of the NVIDIA Grace CPU ISA capabilities. The Grace CPU option uses the Neoverse-V2 core tuning already within the GNU Compiler Collection but there are also a number of architectural differences with Grace compared to the standard Neoverse-V2. So the GCC compiler has already worked on these AArch64 cores, but without any explicit "-mcpu=" option to precisely match the processor capabilities and thus developers/users needing to resort to the more generic -mcpu=neoverse-v2 or rely on manually passing all the compiler feature options.

Thus moving ahead, users of NVIDIA Grace AArch64 CPUs can now use "-mcpu=grace" for simply and easily enabling the Neoverse-V2 features plus I8MM, BF16, SVE2_BITPERM, SVE2_AES, SVE2_SHA3, and SVE2_SM4 architecture features without having to resort to manually enabling any of the extra options.

The GCC "-mcpu=grace" option was merged to GCC 15 Git via [1]this patch for those interested. For those that missed my previous (remote) testing of the Grace CPUs, see the [2]GH200 articles for benchmarks and more.



[1] https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=7fada36c77829a197f63dde0d48ca33139105202

[2] https://www.phoronix.com/search/GH200