SPEC CPU result comparison

15th October 2026 - Normalized report

SPEC2026 FPrate: PGO+LTO -Ofast native - LLVM, ICX and GCC

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

Please note that event though the header of this comparison says -Ofast, benchmark 767.nest_r was additionally compiled with -fno-finite-math-only. Furthermore, benchmarks 737.gmsh_r and 748.flightdm_r were compiled with -fno-fast-math, so a big chunk of -Ofast was in fact disabled. (This might put GCC at a disadvantage compared to LLVM-based compilers because the latter defaults to -fno-math-errno and -fno-trapping-math, in future measurements, we will try to enable these flags.)

AMD EPYC 8635P (zen5 based Sorano)

Run-times

Benchmark LLVM ICX GCC ICX GCC
709.cactus_r 100.00% 36.65% 110.36% 100.00% 301.09%
722.palm_r CE 100.00% 124.72%
731.astcenc_r 100.00% 88.08% 134.23% 100.00% 152.40%
736.ocio_r 100.00% 110.56% 95.56% 100.00% 86.43%
737.gmsh_r 100.00% 90.68% 94.07% 100.00% 103.74%
748.flightdm_r 100.00% 84.45% 83.53% 100.00% 98.90%
749.fotonik3d_r CE 100.00% 103.65%
765.roms_r CE 100.00% 45.84%
766.femflow_r 100.00% 85.71% 95.92% 100.00% 111.90%
767.nest_r 100.00% 98.24% 91.18% 100.00% 92.82%
772.marian_r 100.00% 68.75% 68.75% 100.00% 100.00%
782.lbm_r 100.00% 72.54% 41.19% 100.00% 56.78%
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)
Naturally, the smaller the run-times, the better.

Non-debug executable sizes

Benchmark LLVM ICX GCC ICX GCC
709.cactus_r 100.00% 179.23% 90.49% 100.00% 50.49%
722.palm_r CE 100.00% 78.52%
731.astcenc_r 100.00% 160.59% 94.18% 100.00% 58.65%
736.ocio_r 100.00% 106.17% 82.12% 100.00% 77.34%
737.gmsh_r 100.00% 107.48% 74.71% 100.00% 69.51%
748.flightdm_r 100.00% 124.06% 100.11% 100.00% 80.70%
749.fotonik3d_r CE 100.00% 23.62%
765.roms_r CE 100.00% 62.37%
766.femflow_r 100.00% 106.62% 64.63% 100.00% 60.62%
767.nest_r 100.00% 103.71% 73.39% 100.00% 70.77%
772.marian_r 100.00% 104.03% 83.89% 100.00% 80.64%
782.lbm_r 100.00% 106.78% 40.59% 100.00% 38.02%
Smaller binary sizes are always better.

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

Run-times

Benchmark LLVM ICX GCC ICX GCC
709.cactus_r 100.00% 37.25% 98.04% 100.00% 263.16%
722.palm_r CE 100.00% 126.28%
731.astcenc_r 100.00% 110.80% 114.81% 100.00% 103.62%
736.ocio_r 100.00% 100.50% 90.70% 100.00% 90.25%
737.gmsh_r 100.00% 90.91% 94.50% 100.00% 103.95%
748.flightdm_r 100.00% 91.37% 87.05% 100.00% 95.28%
749.fotonik3d_r CE 100.00% 109.61%
765.roms_r CE 100.00% 77.93%
766.femflow_r 100.00% 100.00% 118.75% 100.00% 118.75%
767.nest_r 100.00% 103.04% 100.51% 100.00% 97.54%
772.marian_r 100.00% 89.66% 106.90% 100.00% 119.23%
782.lbm_r 100.00% 129.11% 109.18% 100.00% 84.56%
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)
Naturally, the smaller the run-times, the better.

Non-debug executable sizes

Benchmark LLVM ICX GCC ICX GCC
709.cactus_r 100.00% 176.20% 92.94% 100.00% 52.74%
722.palm_r CE 100.00% 73.67%
731.astcenc_r 100.00% 160.74% 96.97% 100.00% 60.33%
736.ocio_r 100.00% 104.12% 82.77% 100.00% 79.50%
737.gmsh_r 100.00% 107.96% 75.68% 100.00% 70.09%
748.flightdm_r 100.00% 123.68% 100.14% 100.00% 80.96%
749.fotonik3d_r CE 100.00% 22.34%
765.roms_r CE 100.00% 57.83%
766.femflow_r 100.00% 109.15% 66.76% 100.00% 61.16%
767.nest_r 100.00% 104.92% 74.61% 100.00% 71.12%
772.marian_r 100.00% 103.85% 84.82% 100.00% 81.68%
782.lbm_r 100.00% 154.16% 69.06% 100.00% 44.80%
Smaller binary sizes are always better.