SPEC CPU result comparison

15th October 2026 - Normalized report

SPEC2017 INTrate: 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
500.perlbench_r 100.00% 100.00% 83.16% 100.00% 83.16%
502.gcc_r 100.00% 100.00% 97.47% 100.00% 97.47%
505.mcf_r 100.00% 85.32% 77.45% 100.00% 90.77%
520.omnetpp_r 100.00% 100.00% 88.48% 100.00% 88.48%
523.xalancbmk_r 100.00% 99.19% 87.47% 100.00% 88.19%
525.x264_r 100.00% 101.40% 96.49% 100.00% 95.16%
531.deepsjeng_r 100.00% 100.48% 90.87% 100.00% 90.43%
541.leela_r 100.00% 100.73% 106.36% 100.00% 105.58%
548.exchange2_r CE CE CE
557.xz_r 100.00% 98.96% 95.63% 100.00% 96.64%
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 LLVM FullLTO GCC LLVM FullLTO GCC
500.perlbench_r 100.00% 100.99% 97.42% 100.00% 96.47%
502.gcc_r 100.00% 100.06% 91.57% 100.00% 91.52%
505.mcf_r 100.00% 119.37% 78.90% 100.00% 66.10%
520.omnetpp_r 100.00% 98.23% 81.05% 100.00% 82.51%
523.xalancbmk_r 100.00% 94.86% 78.45% 100.00% 82.70%
525.x264_r 100.00% 104.01% 95.26% 100.00% 91.58%
531.deepsjeng_r 100.00% 96.71% 98.73% 100.00% 102.09%
541.leela_r 100.00% 93.85% 73.52% 100.00% 78.33%
548.exchange2_r CE CE CE
557.xz_r 100.00% 98.36% 93.70% 100.00% 95.26%
Smaller binary sizes are always better.

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

Run-times

Benchmark LLVM LLVM FullLTO GCC LLVM FullLTO GCC
500.perlbench_r 100.00% 96.60% 79.15% 100.00% 81.94%
502.gcc_r 100.00% 100.00% 96.44% 100.00% 96.44%
505.mcf_r 100.00% 86.04% 81.98% 100.00% 95.29%
520.omnetpp_r 100.00% 100.00% 91.80% 100.00% 91.80%
523.xalancbmk_r 100.00% 94.75% 91.60% 100.00% 96.67%
525.x264_r 100.00% 103.27% 84.20% 100.00% 81.53%
531.deepsjeng_r 100.00% 97.70% 103.32% 100.00% 105.74%
541.leela_r 100.00% 99.28% 102.65% 100.00% 103.40%
548.exchange2_r CE CE CE
557.xz_r 100.00% 99.59% 93.47% 100.00% 93.85%
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 LLVM FullLTO GCC LLVM FullLTO GCC
500.perlbench_r 100.00% 100.81% 97.92% 100.00% 97.14%
502.gcc_r 100.00% 99.97% 92.09% 100.00% 92.11%
505.mcf_r 100.00% 122.48% 77.53% 100.00% 63.30%
520.omnetpp_r 100.00% 98.36% 82.37% 100.00% 83.74%
523.xalancbmk_r 100.00% 94.63% 79.79% 100.00% 84.32%
525.x264_r 100.00% 103.91% 97.82% 100.00% 94.14%
531.deepsjeng_r 100.00% 96.63% 98.86% 100.00% 102.31%
541.leela_r 100.00% 92.52% 79.81% 100.00% 86.26%
548.exchange2_r CE CE CE
557.xz_r 100.00% 101.33% 95.34% 100.00% 94.09%
Smaller binary sizes are always better.