SPEC CPU result comparison

15th October 2026 - Normalized report

SPEC2017 INTrate: ThinLTO and LTO native -Ofast - LLVM and GCC

-Ofast -march=native -flto=thin vs. -Ofast -march=native -flto

AMD EPYC 8635P (zen5 based Sorano)

Run-times

Benchmark LLVM ThinLTO LLVM FullLTO GCC LLVM FullLTO GCC
500.perlbench_r 100.00% 98.74% 94.77% 100.00% 95.97%
502.gcc_r 100.00% 99.15% 98.30% 100.00% 99.14%
505.mcf_r 100.00% 81.91% 73.98% 100.00% 90.32%
520.omnetpp_r 100.00% 100.00% 100.00% 100.00% 100.00%
523.xalancbmk_r 100.00% 94.79% 80.42% 100.00% 84.84%
525.x264_r 100.00% 101.59% 97.21% 100.00% 95.69%
531.deepsjeng_r 100.00% 98.66% 92.51% 100.00% 93.77%
541.leela_r 100.00% 100.00% 102.20% 100.00% 102.20%
548.exchange2_r 100.00% 116.23% 89.74% 100.00% 77.21%
557.xz_r 100.00% 99.57% 101.49% 100.00% 101.93%
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)
Naturally, the smaller the run-times, the better.

Non-debug executable sizes

Benchmark LLVM ThinLTO LLVM FullLTO GCC LLVM FullLTO GCC
500.perlbench_r 100.00% 76.11% 87.11% 100.00% 114.46%
502.gcc_r 100.00% 94.15% 95.13% 100.00% 101.04%
505.mcf_r 100.00% 77.05% 64.44% 100.00% 83.63%
520.omnetpp_r 100.00% 98.33% 98.03% 100.00% 99.69%
523.xalancbmk_r 100.00% 91.73% 95.06% 100.00% 103.63%
525.x264_r 100.00% 110.08% 123.10% 100.00% 111.82%
531.deepsjeng_r 100.00% 96.76% 99.03% 100.00% 102.34%
541.leela_r 100.00% 85.60% 91.52% 100.00% 106.91%
548.exchange2_r 100.00% 101.03% 18.81% 100.00% 18.62%
557.xz_r 100.00% 92.46% 104.35% 100.00% 112.86%
Smaller binary sizes are always better.

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

Run-times

Benchmark LLVM ThinLTO LLVM FullLTO GCC LLVM FullLTO GCC
500.perlbench_r 100.00% 106.90% 95.26% 100.00% 89.11%
502.gcc_r 100.00% 98.72% 96.38% 100.00% 97.63%
505.mcf_r 100.00% 76.86% 72.43% 100.00% 94.24%
520.omnetpp_r 100.00% 100.44% 94.25% 100.00% 93.83%
523.xalancbmk_r 100.00% 98.88% 88.79% 100.00% 89.80%
525.x264_r 100.00% 108.33% 95.14% 100.00% 87.82%
531.deepsjeng_r 100.00% 99.52% 95.17% 100.00% 95.63%
541.leela_r 100.00% 99.06% 101.17% 100.00% 102.13%
548.exchange2_r 100.00% 88.60% 82.35% 100.00% 92.95%
557.xz_r 100.00% 99.57% 102.99% 100.00% 103.43%
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)
Naturally, the smaller the run-times, the better.

Non-debug executable sizes

Benchmark LLVM ThinLTO LLVM FullLTO GCC LLVM FullLTO GCC
500.perlbench_r 100.00% 75.66% 87.25% 100.00% 115.33%
502.gcc_r 100.00% 93.51% 95.66% 100.00% 102.31%
505.mcf_r 100.00% 78.41% 63.33% 100.00% 80.77%
520.omnetpp_r 100.00% 98.85% 99.38% 100.00% 100.54%
523.xalancbmk_r 100.00% 91.60% 97.53% 100.00% 106.47%
525.x264_r 100.00% 112.15% 127.33% 100.00% 113.53%
531.deepsjeng_r 100.00% 96.73% 99.07% 100.00% 102.41%
541.leela_r 100.00% 86.03% 104.40% 100.00% 121.35%
548.exchange2_r 100.00% 100.18% 19.38% 100.00% 19.35%
557.xz_r 100.00% 94.29% 106.32% 100.00% 112.77%
Smaller binary sizes are always better.