programming-language-benchmarks

Benchmarks for programming languages

Languages and Implementations*

• C
• C++
• C#
• Go
• Java
• JavaScript
• Lua
• LuaJIT
• PHP
• Python
• Rust
• Zig

Philosophy

• Real-world tests only
  Most developers are not writing code that looks like the 1994 FANNKUCH Benchmark, nor running N-body simulations.

• Use the most popular libraries
  The speed and memory efficiency of programming languages should not be evaluated without considering the ecosystems around them. When benchmarking an HTTPS server in Node.js, most developers will be considering the speed of Express, not http.createServer().

  NEW For web server benchmarks, see andrewmcwattersandco/web-server-benchmarks.

• Use idiomatic code/Do not use micro-optimizations
  One can write inline assembly in C with the non-standard asm keyword or use ffi.cdef from LuaJIT to create C data structures instead of Lua tables, but the tests within this benchmark are designed to reflect typical usage of each respective language.

  Further, some implementations of languages rely on specific hinting to trigger optimizations, like knowing the exact layout of a record ahead of time to avoid growing or shrinking backing memory by powers of two.

  Most developers are not aware of such optimizations and this benchmark makes no attempts to catalog them. Instead, tests are written in a way that first reflects the documentation provided by the language authors, who themselves often do not publicize such optimizations.

Run

./bench [-t tests...] [-l languages...] [-s languages...]

macOS

node

bench runs from /bin/sh. If you have node installed through nvm, you may need to create a symbolic link to the default installation path.

sudo ln -s $(command -v node) /usr/local/bin/node

Tests

• Minimal program
  Test initialization

• Create a record/structure/object
  Test memory allocation

  Note: Creates 8,388,608 records. Last tested against node v24.3.0, 67,108,864 is the maximum array size. All other test languages have their tests calibrated to this number, which is the smallest array length of all of the test languages before one of the tests is terminated abnormally.

  Edit: This test previously created 67,108,864 records.

• Parse JSON
  Test built-in or de facto standard JSON parser

Results

Our findings confirm conventional wisdom that low-level compiled languages are faster than managed code, and that managed code runs faster than interpreted code. The results also show that the performance of a programming language is not solely determined by its implementation, but also by the libraries and frameworks that are used with it.

Physical

Last updated: Thu Aug 21 21:34:46 MST 2025

% ./bench
json
 cpp            mean 1738.3 µs	
 c              mean 3054.0 µs	
 zig            mean 3906.9 µs	
 rust           mean 3692.2 µs	
 go             mean 6968.4 µs	
 java           mean 170049.5 µs	
 cs             mean 35628.1 µs	
 js             mean 33187.1 µs	
 luajit         mean 9938.0 µs	
 php            mean 52403.9 µs	
 lua            mean 49366.6 µs	
 ruby           mean 59341.7 µs	
 python3        mean 30460.8 µs	
minimal
 cpp            mean 1549.0 µs	
 c              mean 1217.1 µs	
 zig            mean 1534.4 µs	
 rust           mean 1550.7 µs	
 go             mean 1542.4 µs	
 java           mean 31884.8 µs	
 cs             mean 22668.2 µs	
 js             mean 28519.2 µs	
 luajit         mean 1602.0 µs	
 php            mean 50319.9 µs	
 lua            mean 1747.0 µs	
 ruby           mean 51489.4 µs	
 python3        mean 21472.0 µs	
record
 cpp            mean 738.2 µs	
 c              mean 1215.0 µs	
 zig            mean 4460.7 µs	
 rust           mean 7152.0 µs	
 go             mean 9709.3 µs	
 java           mean 80830.2 µs	
 cs             mean 378984.8 µs	
 js             mean 626475.8 µs	
 luajit         mean 658569.1 µs	
 php            mean 762029.6 µs	
 lua            mean 1279559.3 µs	
 ruby           mean 1446871.1 µs	
 python3        mean 4236537.6 µs	

Google Sheets

https://docs.google.com/spreadsheets/d/1oBtZTcp4UzRTWnTsBwxr465XAZhVuso33EJINMAKUQo/edit?usp=sharing

Hardware

Physical

MacBook Pro (16-inch, 2021)
Apple M1 Max chip with 10-core CPU, 32-core GPU, and 16-core Neural Engine
64GB unified memory
1TB SSD storage
https://www.apple.com/shop/buy-mac/macbook-pro/16-inch-space-gray-10-core-cpu-32-core-gpu-1tb#

Cloud

GitHub Actions

---

Excluded Programming Languages

_{* We do not benchmark Bash/Shell, TypeScript, or PowerShell. We do not benchmark programming languages predominantly used for a single domain, like Kotlin for Android development, Dart for Flutter, Swift for iOS development, or MicroPython for embedded systems. We are unable to benchmark HTML/CSS, SQL, and assembly. We do not benchmark programming languages that are not widely used in production, like Groovy, Elixir, Scala, Delphi, Lisp, Erlang, Fortran, Ada, F#, OCaml, Gleam, Prolog, COBOL, or Mojo. We do not benchmark programming languages that are no longer supported, updated, or ones which "won't be extended to new workloads" like Visual Basic (.NET) or VBA. We do not benchmark programming languages predominantly used for statistical computing or data analysis, like R or MATLAB. We do not benchmark custom scripting languages like GDScript. We do not benchmark programming languages that are write-ins in the Stack Overflow Developer Survey.}

License

GNU General Public License v2.0