SPEC CPU result comparison

15th October 2026 - Normalized report

SPEC2017 FPrate: PGO+ThinLTO and PGO+LTO native -Ofast - LLVM and GCC

-Ofast -march=native -mtune=native -flto=(thin) and -fprofile-instr-use vs. -profile-use

AMD EPYC 8635P (zen5 based Sorano)

Run-times

Benchmark LLVM LLVM FullLTO GCC LLVM FullLTO GCC
503.bwaves_r CE CE CE
507.cactuBSSN_r CE CE CE
508.namd_r 100.00% 98.90% 110.14% 100.00% 111.36%
510.parest_r 100.00% 100.00% 85.45% 100.00% 85.45%
511.povray_r 100.00% 99.18% 87.67% 100.00% 88.40%
519.lbm_r 100.00% 100.00% 48.62% 100.00% 48.62%
521.wrf_r CE CE CE
526.blender_r 100.00% 98.60% 101.40% 100.00% 102.84%
527.cam4_r CE CE CE
538.imagick_r 100.00% 94.64% 166.07% 100.00% 175.47%
544.nab_r 100.00% 100.00% 101.72% 100.00% 101.72%
549.fotonik3d_r CE CE CE
554.roms_r CE CE CE
LLVM clang version 23.1.3 (/labs/mjambor/git/llvm 0d261d1ca552c95a8f007e061c787ac7132fbcbc), flang version 23.1.3 (/labs/mjambor/git/llvm 0d261d1ca552c95a8f007e061c787ac7132fbcbc)
GCC GCC 17.0.0 20261015 (experimental) (revision 84348add236100e5f69a3e9ec4f7b1038bf2b83d)
In the case of LLVM -fno-finite-math-only was used to compile 511.povray_r (that benchmark only) to avoid a known issue.
Naturally, the smaller the run-times, the better.

Non-debug executable sizes

Benchmark LLVM LLVM FullLTO GCC LLVM FullLTO GCC
503.bwaves_r CE CE CE
507.cactuBSSN_r CE CE CE
508.namd_r 100.00% 86.34% 72.99% 100.00% 84.54%
510.parest_r 100.00% 98.51% 70.92% 100.00% 71.99%
511.povray_r 100.00% 103.47% 85.11% 100.00% 82.25%
519.lbm_r 100.00% 103.79% 61.25% 100.00% 59.02%
521.wrf_r CE CE CE
526.blender_r 100.00% 100.89% 68.78% 100.00% 68.18%
527.cam4_r CE CE CE
538.imagick_r 100.00% 100.86% 78.21% 100.00% 77.55%
544.nab_r 100.00% 98.49% 94.96% 100.00% 96.42%
549.fotonik3d_r CE CE CE
554.roms_r CE CE CE
Smaller binary sizes are always better.

AMD EPYC 75F3 (zen3 based high-frequency Milan)

Run-times

Benchmark LLVM LLVM FullLTO GCC LLVM FullLTO GCC
503.bwaves_r CE CE CE
507.cactuBSSN_r CE CE CE
508.namd_r 100.00% 96.67% 101.43% 100.00% 104.93%
510.parest_r 100.00% 99.61% 100.00% 100.00% 100.39%
511.povray_r 100.00% 99.51% 91.18% 100.00% 91.63%
519.lbm_r 100.00% 100.00% 150.86% 100.00% 150.86%
521.wrf_r CE CE CE
526.blender_r 100.00% 98.27% 107.14% 100.00% 109.03%
527.cam4_r CE CE CE
538.imagick_r 100.00% 95.22% 146.41% 100.00% 153.77%
544.nab_r 100.00% 100.86% 102.87% 100.00% 101.99%
549.fotonik3d_r CE CE CE
554.roms_r CE CE CE
LLVM clang version 23.1.3 (https://github.com/llvm/llvm-project.git 0d261d1ca552c95a8f007e061c787ac7132fbcbc), flang version 23.1.3 (https://github.com/llvm/llvm-project.git 0d261d1ca552c95a8f007e061c787ac7132fbcbc)
GCC GCC 17.0.0 20261015 (experimental) (revision 84348add236100e5f69a3e9ec4f7b1038bf2b83d)
In the case of LLVM -fno-finite-math-only was used to compile 511.povray_r (that benchmark only) to avoid a known issue.
Naturally, the smaller the run-times, the better.

Non-debug executable sizes

Benchmark LLVM LLVM FullLTO GCC LLVM FullLTO GCC
503.bwaves_r CE CE CE
507.cactuBSSN_r CE CE CE
508.namd_r 100.00% 85.56% 75.88% 100.00% 88.68%
510.parest_r 100.00% 98.00% 73.38% 100.00% 74.88%
511.povray_r 100.00% 102.85% 87.13% 100.00% 84.71%
519.lbm_r 100.00% 96.03% 74.92% 100.00% 78.01%
521.wrf_r CE CE CE
526.blender_r 100.00% 100.50% 69.52% 100.00% 69.17%
527.cam4_r CE CE CE
538.imagick_r 100.00% 99.85% 80.53% 100.00% 80.65%
544.nab_r 100.00% 98.44% 96.04% 100.00% 97.56%
549.fotonik3d_r CE CE CE
554.roms_r CE CE CE
Smaller binary sizes are always better.