NVIDIA's Olympus are the ARM64 cores found within the upcoming Vera CPU that will be paired with Rubin. Olympus cores are claimed to be twice as fast as NVIDIA's current CPU cores found in Grace and based on Neoverse-V2. Earlier this year [1]the open-source compilers landed initial support for Olympus while now a proper CPU scheduling model has been upstreamed into LLVM 22.

The initial NVIDIA Olympis compiler enablement that landed earlier this year for the GCC and LLVM/Clang compilers confirmed that it's based on Armv9.2-A capabilities and includes SVE2_BITPERM, RNG, LS64, MEMTAG, PROFILE, FAMINMAX, FP8DOT2, LUT, SVE2_AES, SVE2_SHA3, SVE2_SM4 support.

Earlier this month NVIDIA [2]published a software optimization guide for these custom CPU cores. With that now public, in turn the LLVM 22 compiler has landed its optimized CPU scheduling model so the compiler can make better educated instruction scheduling decisions.

[3]This commit from a NVIDIA engineer was merged last week to LLVM Git for providing this scheduling model for those 88 CPU cores to be found with NVIDIA's upcoming Vera-Rubin servers.



[1] https://www.phoronix.com/news/NVIDIA-Olympus-GCC-Clang

[2] https://docs.nvidia.com/olympus-cpu-core-software-optimization-guide-dp12531-001v0-7.pdf

[3] https://github.com/llvm/llvm-project/commit/6299dbca5b3d