SPEC CPU result comparison

15th October 2026 - Normalized report

SPEC2017 INTrate: Trunk native -Ofast - standard vs LTO vs LTO+PGO

-Ofast -march=native -mtune=native

AMD EPYC 8635P (zen5 based Sorano)

Run-times

Benchmark Normal LTO LTO+PGO LTO LTO+PGO
500.perlbench_r 100.00% 97.21% 85.84% 100.00% 88.30%
502.gcc_r 100.00% 97.47% 89.26% 100.00% 91.58%
505.mcf_r 100.00% 85.25% 85.25% 100.00% 100.00%
520.omnetpp_r 100.00% 92.84% 83.37% 100.00% 89.80%
523.xalancbmk_r 100.00% 95.54% 107.18% 100.00% 112.18%
525.x264_r 100.00% 93.13% 104.96% 100.00% 112.70%
531.deepsjeng_r 100.00% 82.38% 90.00% 100.00% 109.25%
541.leela_r 100.00% 92.48% 96.24% 100.00% 104.07%
548.exchange2_r 100.00% 96.44% 98.93% 100.00% 102.58%
557.xz_r 100.00% 98.96% 95.63% 100.00% 96.64%
The compiler used was GCC 17.0.0 20261015 (experimental) (revision 84348add236100e5f69a3e9ec4f7b1038bf2b83d).
Naturally, the smaller the run-times, the better.
See also continuous gcc options benchmarking at https://lnt.opensuse.org/db_default/v4/SPEC/spec_report/options.

Non-debug executable sizes

Benchmark Normal LTO LTO+PGO LTO LTO+PGO
500.perlbench_r 100.00% 111.02% 75.93% 100.00% 68.40%
502.gcc_r 100.00% 103.85% 90.39% 100.00% 87.04%
505.mcf_r 100.00% 78.00% 57.77% 100.00% 74.06%
520.omnetpp_r 100.00% 82.91% 59.15% 100.00% 71.34%
523.xalancbmk_r 100.00% 72.15% 46.53% 100.00% 64.50%
525.x264_r 100.00% 124.00% 88.27% 100.00% 71.19%
531.deepsjeng_r 100.00% 99.12% 99.17% 100.00% 100.06%
541.leela_r 100.00% 56.80% 40.45% 100.00% 71.22%
548.exchange2_r 100.00% 87.90% 49.21% 100.00% 55.99%
557.xz_r 100.00% 71.30% 53.91% 100.00% 75.61%
Smaller binary sizes are always better.

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

Run-times

Benchmark Normal LTO LTO+PGO LTO LTO+PGO
500.perlbench_r 100.00% 99.10% 83.41% 100.00% 84.16%
502.gcc_r 100.00% 97.00% 86.94% 100.00% 89.62%
505.mcf_r 100.00% 83.53% 84.45% 100.00% 101.11%
520.omnetpp_r 100.00% 91.61% 86.67% 100.00% 94.60%
523.xalancbmk_r 100.00% 92.52% 101.87% 100.00% 110.10%
525.x264_r 100.00% 93.20% 105.10% 100.00% 112.77%
531.deepsjeng_r 100.00% 88.34% 90.81% 100.00% 102.79%
541.leela_r 100.00% 88.34% 87.12% 100.00% 98.61%
548.exchange2_r 100.00% 95.73% 96.58% 100.00% 100.89%
557.xz_r 100.00% 99.18% 94.24% 100.00% 95.02%
The compiler used was GCC 17.0.0 20261015 (experimental) (revision 84348add236100e5f69a3e9ec4f7b1038bf2b83d).
Naturally, the smaller the run-times, the better.
See also continuous gcc options benchmarking at https://lnt.opensuse.org/db_default/v4/SPEC/spec_report/options.

Non-debug executable sizes

Benchmark Normal LTO LTO+PGO LTO LTO+PGO
500.perlbench_r 100.00% 111.03% 76.24% 100.00% 68.67%
502.gcc_r 100.00% 104.05% 90.75% 100.00% 87.22%
505.mcf_r 100.00% 80.14% 61.21% 100.00% 76.37%
520.omnetpp_r 100.00% 82.86% 59.35% 100.00% 71.63%
523.xalancbmk_r 100.00% 71.07% 45.49% 100.00% 64.01%
525.x264_r 100.00% 124.63% 87.79% 100.00% 70.44%
531.deepsjeng_r 100.00% 99.09% 99.24% 100.00% 100.15%
541.leela_r 100.00% 56.69% 40.42% 100.00% 71.31%
548.exchange2_r 100.00% 87.25% 48.18% 100.00% 55.21%
557.xz_r 100.00% 71.18% 53.46% 100.00% 75.10%
Smaller binary sizes are always better.