Performance

This page shows recorded timings. None of the numbers are typed by hand: when the site is built, the six judge problems are read from examples/15_algorithms/measurements.json and the collatz table from the README.

scripts/refresh.py re-measures and reports what has drifted past a tolerance, and --write records it.

Collatz kernel compared with other compilers

This is the collatz kernel from the README. It was run on one machine, ten runs each in a fresh process, and the table shows the kernel's wall time (mean ± standard deviation).

Of the PPy rows, ppy build --unsafe is the wrap-semantics artifact. ppy run keeps Python-integer semantics: overflow is guarded and falls back to arbitrary precision.

compiler	kernel	integer semantics
PPy ppy build --unsafe --host-cpu	28.8 ± 0.5 ms	64-bit, wraps on overflow (this machine's instruction set)
C (clang -O3 -march=native)	30.8 ± 0.7 ms	64-bit, wraps on overflow
PPy ppy build --unsafe	30.8 ± 0.4 ms	64-bit, wraps on overflow
Numba @njit	32.5 ± 0.9 ms	64-bit, wraps on overflow
C (clang -O3)	32.8 ± 0.5 ms	64-bit, wraps on overflow
PPy ppy run	41.7 ± 1.8 ms	Python ints: guarded, falls back to arbitrary precision
C (gcc -O3)	41.9 ± 0.9 ms	64-bit, wraps on overflow
Codon -release	42.1 ± 12.1 ms	64-bit, wraps on overflow
PyPy 3.11	52.8 ± 0.5 ms	Python ints
Cython (cdef long long)	61.6 ± 0.8 ms	64-bit, wraps on overflow
mypyc	71.6 ± 1.4 ms	Python ints
Nuitka	714.7 ± 12.8 ms	Python ints, no type specialization
CPython 3.14	1068.8 ± 5.5 ms	Python ints

Six judge problems, whole process

These are wall times of the whole process, including input reading and interpreter startup. Nothing inside the programs is instrumented. Each problem reads its input from standard input.

problem	CPython	ppy run	ppy build --unsafe	--standalone	C (gcc -O3)	C (clang -O3)
N-Queens 15a_nqueens	142.0 ± 12.5 ms	123.6 ± 144.1 ms	48.7 ± 14.9 ms	5.7 ± 0.2 ms	4.8 ± 0.1 ms	5.4 ± 0.1 ms
Shortest paths 15b_dijkstra	4858.5 ± 72.8 ms	3659.1 ± 238.0 ms	4426.9 ± 53.7 ms	111.7 ± 1.2 ms	142.2 ± 2.5 ms	134.3 ± 2.4 ms
Substring search 15c_kmp	292.4 ± 12.8 ms	145.5 ± 160.1 ms	53.3 ± 1.3 ms	—	9.3 ± 0.1 ms	9.0 ± 0.4 ms
Range sums 15d_segment_tree	1571.7 ± 26.0 ms	1288.3 ± 157.9 ms	1524.4 ± 24.2 ms	30.1 ± 0.8 ms	50.8 ± 0.7 ms	50.9 ± 0.9 ms
Longest increasing subsequence 15e_lis	541.6 ± 14.0 ms	198.6 ± 152.9 ms	111.5 ± 1.6 ms	41.5 ± 0.6 ms	57.7 ± 0.4 ms	53.9 ± 0.4 ms
Counting inversions 15f_input	603.7 ± 11.2 ms	189.3 ± 166.5 ms	98.4 ± 3.1 ms	38.6 ± 0.5 ms	43.5 ± 0.8 ms	44.2 ± 0.4 ms

Recorded on Intel(R) Core(TM) Ultra 9 386H, 16 cores, CPython 3.14.5, Linux-6.6.114.1-microsoft-standard-WSL2-x86_64-with-glibc2.39, at 2026-09-14T04:47:22.

The ppy run column includes the first run's build into the cache. That is why its deviation is wide. From the second run on, it is the launcher alone.

--standalone is a native executable with no interpreter inside.