SPEC CPU result comparison

15th October 2026 - Normalized report

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

-Ofast -native -flto -fprofile-instr-use vs -Ofast -march=native -flto -fprofile-use

AMD EPYC 8635P (zen5 based Sorano)

Run-times

Benchmark LLVM ICX GCC ICX GCC
503.bwaves_r CE 100.00% 92.75%
507.cactuBSSN_r CE 100.00% 105.98%
508.namd_r 100.00% 103.60% 111.36% 100.00% 107.49%
510.parest_r 100.00% 88.73% 85.45% 100.00% 96.31%
511.povray_r 100.00% 92.82% 88.40% 100.00% 95.24%
519.lbm_r 100.00% 80.88% 48.62% 100.00% 60.11%
521.wrf_r CE 100.00% 106.67%
526.blender_r 100.00% 90.54% 102.84% 100.00% 113.58%
527.cam4_r CE 100.00% 102.14%
538.imagick_r 100.00% 159.12% 175.47% 100.00% 110.28%
544.nab_r 100.00% 86.49% 101.72% 100.00% 117.61%
549.fotonik3d_r CE 100.00% 94.33%
554.roms_r CE 100.00% 41.45%
LLVM clang version 23.1.3 (/labs/mjambor/git/llvm 0d261d1ca552c95a8f007e061c787ac7132fbcbc), flang version 23.1.3 (/labs/mjambor/git/llvm 0d261d1ca552c95a8f007e061c787ac7132fbcbc)
ICX Intel(R) oneAPI DPC++/C++ Compiler 2026.1.1 (2026.1.1.20260724), ifx (IFX) 2026.1.1 20260724
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 ICX GCC ICX GCC
503.bwaves_r CE 100.00% 4.74%
507.cactuBSSN_r CE 100.00% 41.57%
508.namd_r 100.00% 157.25% 84.54% 100.00% 53.76%
510.parest_r 100.00% 124.47% 71.99% 100.00% 57.84%
511.povray_r 100.00% 160.85% 82.25% 100.00% 51.13%
519.lbm_r 100.00% 141.49% 59.02% 100.00% 41.71%
521.wrf_r CE 100.00% 84.20%
526.blender_r 100.00% 86.82% 68.18% 100.00% 78.53%
527.cam4_r CE 100.00% 86.94%
538.imagick_r 100.00% 220.04% 77.55% 100.00% 35.24%
544.nab_r 100.00% 167.05% 96.42% 100.00% 57.72%
549.fotonik3d_r CE 100.00% 23.65%
554.roms_r CE 100.00% 52.12%
Smaller binary sizes are always better.

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

Run-times

Benchmark LLVM ICX GCC ICX GCC
503.bwaves_r CE 100.00% 100.75%
507.cactuBSSN_r CE 100.00% 101.70%
508.namd_r 100.00% 104.93% 104.93% 100.00% 100.00%
510.parest_r 100.00% 100.39% 100.39% 100.00% 100.00%
511.povray_r 100.00% 95.57% 91.63% 100.00% 95.88%
519.lbm_r 100.00% 147.84% 150.86% 100.00% 102.04%
521.wrf_r CE 100.00% 110.63%
526.blender_r 100.00% 88.33% 109.03% 100.00% 123.44%
527.cam4_r CE 100.00% 105.44%
538.imagick_r 100.00% 111.06% 153.77% 100.00% 138.46%
544.nab_r 100.00% 101.99% 101.99% 100.00% 100.00%
549.fotonik3d_r CE 100.00% 100.46%
554.roms_r CE 100.00% 70.82%
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)
ICX Intel(R) oneAPI DPC++/C++ Compiler 2026.1.1 (2026.1.1.20260724), ifx (IFX) 2026.1.1 20260724
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 ICX GCC ICX GCC
503.bwaves_r CE 100.00% 4.76%
507.cactuBSSN_r CE 100.00% 37.09%
508.namd_r 100.00% 154.23% 88.68% 100.00% 57.50%
510.parest_r 100.00% 127.28% 74.88% 100.00% 58.84%
511.povray_r 100.00% 163.21% 84.71% 100.00% 51.90%
519.lbm_r 100.00% 155.17% 78.01% 100.00% 50.28%
521.wrf_r CE 100.00% 79.68%
526.blender_r 100.00% 86.53% 69.17% 100.00% 79.94%
527.cam4_r CE 100.00% 86.17%
538.imagick_r 100.00% 233.51% 80.65% 100.00% 34.54%
544.nab_r 100.00% 168.37% 97.56% 100.00% 57.95%
549.fotonik3d_r CE 100.00% 22.33%
554.roms_r CE 100.00% 49.04%
Smaller binary sizes are always better.