Speed Tests

When I learn a new programming language, I always implement the Münchausen numbers problem in the given language. The problem is simple but it includes a lot of computations, thus it gives an idea of the execution speed of a language.

Münchausen numbers

A Münchausen number is a number equal to the sum of its digits raised to each digit's power.

For instance, 3435 is a Münchausen number because 3³+4⁴+3³+5⁵ = 3435.

0⁰ is not well-defined, thus we'll consider 0⁰=0. In this case there are four Münchausen numbers: 0, 1, 3435, and 438579088.

Exercise

Write a program that finds all the Münchausen numbers. We know that the largest Münchausen number is less than 440 million.

Updates

Dates are in yyyy-mm-dd format.

2024-October: Java and Kotlin runtimes were updated. PHP was added. Clojure was added. Python was updated to version 3.12. JavaScript got a faster version. Go got a faster version. Nim was updated to version 2.2. Parallel Go was added. Crystal and Odin were re-tested. Mojo was added. Scala 3 was added.

Implementations

In the implementations I tried to use the same (simple) algorithm in order to make the comparisons as fair as possible.

All the tests were run on my home desktop machine (Intel Core i7-4771 CPU @ 3.50GHz with 8 CPU cores) using Manjaro Linux. Execution times are wall-clock times and they are measured with hyperfine (warmup runs: 1, benchmarked runs: 2).

The following implementations were received in the form of pull requests:

• Clojure, Common LISP, Crystal, D, FASM, Fortran, Haskell, JavaScript, Lua, Mojo, NASM, OCaml, Pascal, Perl, PHP, Python 3 with Numba, Racket, Ruby, Scala 3, Scheme, Swift, Toit, V, Zig

Thanks for the contributions!

If you know how to make something faster, let me know!

Languages are listed in alphabetical order.

The size of the EXE files can be further reduced with the command strip -s. If it's applicable, then the stripped EXE size is also shown in the table.

Below, you can find single-threaded implemetations. We also have some multi-threaded implementations, see here.

---

C

• gcc (GCC) 13.2.1 20230801
• clang version 16.0.6
• Benchmark date: 2024-02-05 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
gcc -O3 main.c -o main -lm	3.893 ± 0.01	15,560	14,408
gcc -O2 main.c -o main -lm	3.892 ± 0.001	15,560	14,408
clang -O3 main.c -o main -lm	2.684 ± 0.013	15,528	14,416
clang -O2 main.c -o main -lm	2.672 ± 0.001	15,528	14,416

Notes:

• No real difference between the switches -O2 and -O3. It's enough to use -O2.
• clang is better in this case

see source

C#

• dotnet 8.0.100
• Benchmark date: 2024-02-07 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	--
dotnet publish -o dist -c Release	5.614 ± 0.097	603,488	--

Note: almost the same performance as Java.

see source

C++

• g++ (GCC) 13.2.1 20230801
• clang version 16.0.6
• Benchmark date: 2024-02-05 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
g++ -O3 --std=c++2a main.cpp -o main	3.865 ± 0.01	15,936	14,432
g++ -O2 --std=c++2a main.cpp -o main	3.849 ± 0.012	15,936	14,432
clang++ -O3 --std=c++2a main.cpp -o main	2.913 ± 0.01	15,904	14,440
clang++ -O2 --std=c++2a main.cpp -o main	2.827 ± 0.015	15,904	14,440

Notes:

• No big difference between the switches -O2 and -O3. Using -O2 is even better.
• clang is better in this case

see source

Clojure

• Clojure CLI version 1.12.0.1479
• Benchmark date: 2024-10-08 [yyyy-mm-dd]

Execution	Runtime (sec)	compiled / transpiled output size (bytes)	--
clj -M -m main	5.631 ± 0.112	--	--
mkdir classes && java -cp `clj -Spath` main	5.339 ± 0.101	--	--

see source

Notes:

• A bit slower than Java.

Codon

• codon 0.15.5
• Benchmark date: 2023-04-02 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
codon build -release main.py	5.369 ± 0.006	28,400	26,864

Notes:

• Codon is a high-performance Python compiler that compiles Python code to native machine code without any runtime overhead.
• It's a bit faster than C#!
• The code is unchanged Python code. No type annotations are needed.

See https://github.com/exaloop/codon for more information about this compiler.

see source

Common LISP

• GNU CLISP 2.49.93+ (2018-02-18) (built on arojas [135.181.138.48])
• SBCL 2.2.5
• Benchmark date: 2022-09-02 [yyyy-mm-dd]

Execution	Runtime (sec)	--	--
clisp -C main.cl	321.049 ± 0.484	--	--
sbcl --script main.cl	6.828 ± 0.003	--	--

Notes:

• clisp is very slow. Almost as slow as Python. And without the -C switch, it's ten times slower.
• With sbcl, you can get excellent performance.

see source

Crystal

• Crystal 1.13.2 (2024-09-08); LLVM: 18.1.8; Default target: x86_64-pc-linux-gnu
• Benchmark date: 2024-10-13 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
crystal build --release main.cr	4.237 ± 0.077	807,432	273,424

Notes:

• The runtime is very good, similar to Go.
• The source code is almost identical to the Ruby source code.
• The build time is also good. In a previous version (2022) it was painfully slow.

See https://crystal-lang.org for more info about this language.

see source

D

• DMD64 D Compiler v2.100.0
• LDC - the LLVM D compiler (1.29.0)
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
dmd -release -O main.d	9.987 ± 0.045	993,816	712,504
ldc2 -release -O main.d	3.089 ± 0.008	34,584	23,008

Notes:

• the runtime is comparable to C/C++
• the official compiler dmd is slow
• ldc2 is the best in this case

see source

Dart

• Dart SDK version: 2.17.6 (stable) (Tue Jul 12 12:54:37 2022 +0200) on "linux_x64"
• Node.js v18.6.0
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Execution	Runtime (sec)	compiled / transpiled output size (bytes)	--
dart main.dart	23.909 ± 0.581	--	--
dart compile js main.dart -O2 -m -o main.js && node main.js	10.509 ± 0.032	31,684	--
dart compile exe main.dart -o main && ./main	8.377 ± 0.009	5,925,856	--

(*): in the first case, the Dart code is executed as a script

Notes:

• If you execute it as a script (JIT), it's slow.
• If you compile to native code (AOT), it's fast (though slower than Java/C#).
• stripping damaged the EXE file

see source

Elixir

• Erlang/OTP 24 [erts-12.3] [source] [64-bit] [smp:8:8] [ds:8:8:10] [async-threads:1] [jit]; Elixir 1.13.2 (compiled with Erlang/OTP 24)
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Execution	Runtime (sec)	--	--
elixir main.exs	227.963 ± 0.543	--	--
elixirc munchausen.ex && elixir caller.exs	217.528 ± 0.762	--	--

Notes:

• Elixir doesn't excel in CPU-intensive tasks.
• In the second case, the modules were compiled to .beam files. However, it didn't make the program much faster. The difference is very small.

see source

FASM

• flat assembler version 1.73.30
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
# FASM x64, see v1 in Makefile	15.792 ± 0.018	532	532
# FASM x86, see v2 in Makefile	15.207 ± 0.023	444	444

Note: no difference between the 32-bit and 64-bit versions.

See https://en.wikipedia.org/wiki/FASM for more info about FASM.

see source

Fortran

• GNU Fortran (GCC) 12.1.0
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
gfortran -O2 main.f08 -o main	3.884 ± 0.054	21,016	14,456

Note: its speed is comparable to C.

see source

Go

• go version go1.23.1 linux/amd64
• Benchmark date: 2024-10-08 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
# using int, see v1 in Makefile	4.122 ± 0.034	2,137,820	1,391,192
# using uint and uint32, see v2 in Makefile	3.5 ± 0.045	2,137,756	1,391,192

Notes:

• The speed is between C and Java (slower than C, faster than Java).
• Using uint and uint32, you can get better performance.
• The EXE is quite big.

see source

Haskell

• The Glorious Glasgow Haskell Compilation System, version 8.10.7
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
# basic, see v1 in Makefile	93.816 ± 0.043	3,175,704	754,008
# optimized, see v2 in Makefile	3.517 ± 0.009	6,324,936	3,183,648

Notes:

• The performance of the optimized version is comparable to C.
• However, when you compile the optimized version for the first time, the compilation is very slow.

see source

Java

• openjdk version "21.0.4" 2024-07-16
• Benchmark date: 2024-10-08 [yyyy-mm-dd]

Execution	Runtime (sec)	Binary size (bytes)	--
javac Main.java && java Main	5.003 ± 0.002	1,027	--

(*): the binary size is the size of the .class file

Note: very good performance.

see source

JavaScript

• Node.js v22.8.0
• Benchmark date: 2024-10-08 [yyyy-mm-dd]

Execution	Runtime (sec)	--	--
node main1.js	17.789 ± 0.009	--	--
node main2.js	6.819 ± 0.001	--	--

Notes:

• main1.js is a straightforward implementation
• main2.js is an improved implementation, using a more optimal cache array size

see source

Julia

• julia version 1.10.0
• Benchmark date: 2024-02-07 [yyyy-mm-dd]

Execution	Runtime (sec)	--	--
julia --startup=no main.jl	3.656 ± 0.006	--	--

Note: excellent performance, almost like C.

See https://julialang.org for more info about this language.

see source

Kotlin

• Kotlin version 2.0.20-release-360 (JRE 21.0.4+7)
• openjdk version "21.0.4" 2024-07-16
• Benchmark date: 2024-10-08 [yyyy-mm-dd]

Execution	Runtime (sec)	JAR size (bytes)	--
kotlinc main.kt -include-runtime -d main.jar && java -jar main.jar	5.092 ± 0.004	4,826,841	--

Note: same performance as Java.

see source

Lua

• Lua 5.4.4 Copyright (C) 1994-2022 Lua.org, PUC-Rio
• LuaJIT 2.1.0-beta3 -- Copyright (C) 2005-2022 Mike Pall. https://luajit.org/
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Compilation	Runtime (sec)	--	--
lua main.lua	118.23 ± 1.834	--	--
luajit main.lua	19.694 ± 0.009	--	--

Notes:

• LuaJIT is a Just-In-Time Compiler for Lua. The language evolved and it contains an integer division operator (//), but LuaJIT doesn't understand it.
• The Lua code ran much faster than the Python 3 (CPython) code.
• LuaJIT is fast. Its performance is similar to PyPy3 (even a little bit faster).

see source

Mojo

• mojo 24.5.0 (e8aacb95)
• Benchmark date: 2024-10-18 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
magic run mojo build -o main main.mojo	3.315 ± 0.148	1,160,736	302,952

Notes:

• The execution speed is very impressive. It's even faster than C with gcc.
• The source code is very similar to Python, though not completely identical.

See https://www.modular.com/mojo for more info about Mojo.

see source

NASM

• NASM version 2.15.05 compiled on Sep 24 2020
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
# NASM x86, see v2 in Makefile	15.19 ± 0.012	9,228	8,428
# NASM x64, see v1 in Makefile	15.186 ± 0.034	9,656	8,552

Note: no difference between the 32-bit and 64-bit versions.

See https://en.wikipedia.org/wiki/Netwide_Assembler for more info about NASM.

see source

Nim Tests #1

• Nim Compiler Version 2.2.0 [Linux: amd64]
• gcc (GCC) 14.2.1 20240910
• clang version 18.1.8
• Benchmark date: 2024-10-08 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
nim c -d:release main.nim	3.793 ± 0.021	73,440	63,968
nim c --cc:clang -d:release main.nim	3.614 ± 0.042	57,064	47,552
nim c --cc:clang -d:danger main.nim	3.344 ± 0.019	42,456	35,096
nim c -d:danger main.nim	3.092 ± 0.01	54,584	47,360

(*): if --cc:clang is missing, then the default gcc was used

Notes:

• excellent performance, comparable to C
• danger mode gave some performance boost
• In release mode, there isn't much difference between gcc and clang.
• In danger mode, gcc performs better.

see source

Nim Tests #2

• Nim Compiler Version 2.2.0 [Linux: amd64]
• gcc (GCC) 14.2.1 20240910
• clang version 18.1.8
• Benchmark date: 2024-10-08 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
# using int32, see v3 in Makefile	3.685 ± 0.018	57,064	47,552
# using int64, see v2 in Makefile	3.599 ± 0.009	57,064	47,552
# using int, see v1 in Makefile	3.586 ± 0.004	57,064	47,552
# using uint64, see v5 in Makefile	3.346 ± 0.018	57,112	47,552
# using uint32, see v4 in Makefile	2.904 ± 0.013	57,112	47,552

Here, we used the compiler options --cc:clang -d:release everywhere and tested the different integer data types.

Notes:

• In Nim, the size of int is platform-dependent, i.e. it's 64-bit long on a 64 bit system. Thus, on a 64 bit system, there is no difference between using int and int64 (that is, v1 and v2 are equivalent).
• There's a small difference between int / int64 (signed) and uint64 (unsigned). uint64 is a bit faster.
• int32 (v3) was slower than int64, and uint32 (v4) was faster than uint64 (v5)
• To sum up: you can use int, but if you need some performance gain, try uint32 too. Avoid int32.

see source

OCaml

• ocamlopt 5.1.0
• Benchmark date: 2024-02-05 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
ocamlopt -unsafe -O3 -o main -rounds 10 main.ml	8.18 ± 0.001	1,086,200	902,232

see source

Odin

• odin version dev-2024-10-nightly
• Benchmark date: 2024-10-13 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
odin build . -no-bounds-check -disable-assert -o:speed	3.536 ± 0.338	151,704	145,616

See https://odin-lang.org for more info about this language.

Notes:

• Very good performance, comparable to C.
• A previous version (2022) was slower, so Odin has improved a lot.

see source

Pascal

• Free Pascal Compiler version 3.2.2 [2022/03/02] for x86_64
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
fpc -O3 main.pas	17.454 ± 0.02	531,024	531,024

Notes:

• Three times slower than Java.
• Strangely, strip didn't make the EXE any smaller.

see source

Perl

• This is perl 5, version 38, subversion 1 (v5.38.1) built for x86_64-linux-thread-multi
• Benchmark date: 2024-02-05 [yyyy-mm-dd]

Execution	Runtime (sec)	--	--
perl main.pl	494.71 ± 4.649	--	--
perl -Minteger main.pl	423.805 ± 2.471	--	--

see source

Notes:

• This is the slowest solution. It's even slower than Python.

PHP

• PHP 8.3.11 (cli) (built: Aug 28 2024 18:04:39) (NTS)
• Benchmark date: 2024-10-08 [yyyy-mm-dd]

Execution	Runtime (sec)	--	--
php main.php	181.492 ± 0.536	--	--

see source

Notes:

• Faster than Python 3

Python 3

• Python 3.12.5
• Python 3.10.14 (39dc8d3c85a7, Aug 30 2024, 08:27:45) [PyPy 7.3.17 with GCC 14.2.1 20240805]
• Benchmark date: 2024-10-08 [yyyy-mm-dd]

Execution	Runtime (sec)	--	--
python3 main.py	257.15 ± 1.472	--	--
pypy3 main.py	20.05 ± 0.019	--	--

Notes:

• Python 3.11 was 233 seconds. Version 3.12 got 20+ seconds slower :(
• PyPy3 is fast and comparable to LuaJIT

see source

Python 3 with mypyc

• Python 3.10.5
• mypy 0.971 (compiled: no)
• Benchmark date: 2022-08-12 [yyyy-mm-dd]

Execution	Runtime (sec)	.so (bytes)	stripped .so (bytes)
mypyc main.py && ./start_v3.sh	80.481 ± 0.574	183,992	92,824

Notes:

• mypyc can compile a module. This way, the program can be 4 to 5 times faster.

see source

Python 3 with Nim

• Python 3.10.5
• Nim Compiler Version 1.6.6 [Linux: amd64]
• Benchmark date: 2022-08-13 [yyyy-mm-dd]

Execution	Runtime (sec)	--	--
./start_v1.sh	46.772 ± 0.203	--	--

Notes:

• When you start it for the first time, it'll compile the Nim code as a shared library. Thus the first run may be slower.
• The real work is done in Nim. The Nim code is compiled as a shared library. The Python code just calls a function implemented in Nim.

see source

Python 3 with Numba

• Python 3.11.6
• numba 0.58.1
• numpy 1.26.3
• Benchmark date: 2024-02-07 [yyyy-mm-dd]

Execution	Runtime (sec)	--	--
python3 main.py	5.526 ± 0.435	--	--

Notes:

• Numba is an open source JIT compiler that translates a subset of Python and NumPy code into fast machine code. More info here: https://numba.pydata.org
• The performance is excellent (similar to Java's).
• Almost equivalent to the original Python 3 source code.
• This implementation uses a numpy array for the cache.

see source

Python 3 with Rust

• Python 3.10.5
• rustc 1.62.1 (e092d0b6b 2022-07-16)
• Benchmark date: 2022-08-12 [yyyy-mm-dd]

Compilation	Runtime (sec)	--	--
# see v1 in Makefile && ./start_v1.sh	40.263 ± 1.152	--	--

Notes:

• The real work is done in Rust. The Rust code is compiled as a shared library. The Python code just calls a function implemented in Rust.
• The Rust code uses pyo3. Compilation is done with maturin.

see source

Racket

• Welcome to Racket v8.5 [cs].
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Execution	Runtime (sec)	--	--
racket main.rkt	105.218 ± 0.312	--	--

See https://racket-lang.org for more info about this language.

see source

Ruby

• ruby 3.0.4p208 (2022-04-12 revision 3fa771dded) [x86_64-linux]
• Benchmark date: 2022-07-30 [yyyy-mm-dd]

Execution	Runtime (sec)	--	--
ruby main.rb	199.632 ± 3.2	--	--
ruby --jit main.rb	75.863 ± 1.174	--	--

Notes:

• much faster than Python 3
• When run in JIT mode, the performance is the same as Python's mypyc variant (where mypyc compiles a module).
• PyPy3 is 3-4 times faster than the JIT mode.

see source

Rust

• rustc 1.62.1 (e092d0b6b 2022-07-16)
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
cargo build --release	2.936 ± 0.078	3,839,048	317,752

Notes:

• excellent performance (comparable to C/C++)
• The EXE is very big (almost 4 MB). However, if you strip the EXE, the size becomes acceptable.

see source

Scala 3

• Scala compiler version 3.6.0 -- Copyright 2002-2024, LAMP/EPFL
• openjdk version "21.0.4" 2024-07-16
• Benchmark date: 2024-10-20 [yyyy-mm-dd]

Execution	Runtime (sec)	JAR size (bytes)	--
scalac main.scala -d main.jar && scala main.jar	5.378 ± 0.015	5,782	--

Notes:

• a bit slower than Java, a bit faster than C# (.NET 8)
• same performance as Clojure

see source

Scheme

• chez 9.5.8
• guile (GNU Guile) 2.2.7
• gambitc v4.9.4
• stalin 0.11
• Benchmark date: 2022-09-18 [yyyy-mm-dd]

Execution	Runtime (sec)	EXE (bytes)	--
guile -s main.scm	148.423 ± 1.773	--	--
chez --compile-imported-libraries --optimize-level 3 -q --script main.scm	69.826 ± 0.387	--	--
gambitc -:debug=pqQ0 -exe -cc-options '-O3' main.scm && ./main	21.718 ± 0.229	9,098,392	--
stalin -architecture amd64 -s -On -Ot -Ob -Om -Or -dC -dH -dP\ && ./main	4.599 ± 0.017	25,472	--
stalin -architecture amd64 -s -On -Ot -Ob -Om -Or -dC -dH -dP\ && ./main	4.012 ± 0.014	25,512	--

Note: stalin's performance is close to C.

see source

Swift

• Swift version 5.9.2 (swift-5.9.2-RELEASE)
• Benchmark date: 2024-02-05 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
swiftc -Ounchecked main.swift	3.335 ± 0.004	15,832	11,984

Note: the performance is similar to C++.

see source

Toit

• Toit version: v2.0.0-alpha.74
• Benchmark date: 2023-04-02 [yyyy-mm-dd]

Execution	Runtime (sec)	EXE (bytes)	--
toit.run main.toit	120.263 ± 0.069	--	--
toit.compile -O2 -o main main.toit && ./main	118.63 ± 0.774	1,254,784	1,254,784

Notes:

• The runtime of toit.run and toit.compile is the same. I'm not sure, but I think toit.run compiles to a temp. folder and starts the program from there.
• toit.compile must produce a stripped EXE. Stripping the EXE explicitly didn't change the file size.
• see https://toitlang.org for more info about this language

see source

V

• V 0.3.0 82db1e4
• Benchmark date: 2022-07-28 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
v -prod main.v	4.056 ± 0.004	209,392	187,728
v -cc clang -prod main.v	3.936 ± 0.018	212,720	191,736

By default, it uses GCC.

Notes:

• its speed is comparable to C
• see https://vlang.io for more info about this language

see source

Zig

• zig 0.11.0
• Benchmark date: 2024-02-07 [yyyy-mm-dd]

Compilation	Runtime (sec)	EXE (bytes)	stripped EXE (bytes)
zig build-exe -OReleaseFast src/main.zig	2.975 ± 0.037	1,721,168	170,968

Notes:

• excellent performance (comparable to C/C++)
• see https://ziglang.org for more info about this language

see source