In this presentation we share our experiences correlating static performance predictions made by LLVM-MCA with measured runtime performance metrics for the AArch64 NVIDIA Grace CPU. Accurate static predictions are crucial for LLVM-MCA users (such as compiler engineers) investigating performance bottlenecks or evaluating different assembly code sequences, and for other tools like the SIMD superoptimiser “Minotaur” that rely on LLVM-MCA for performance estimations. The performance predictions should capture the trends of the results obtained on hardware. We show correlation results for several benchmarks and discuss the reasons why LLVM-MCA's predictions are sometimes off and how they could be improved.