If i'm understanding the repository correctly, it looks like each language reads from a file, does some I/O printing to console, then computes the value, then some more console printing and exits.
In my opinion, the comparisons could be better if the file I/O and console printing were removed.
Author here — yep, you understood the repository correctly.
My Python version is a good example of the structure: read rounds.txt, run the loop, print the result, exit. I’m timing the whole program with hyperfine.
I agree that for a “pure compute” microbenchmark you could remove file I/O and console output. I kept them mainly because:
- It gives every language the same simple interface (same input, same output) and acts as a basic correctness/sanity check.
- The benchmark runs 1 billion iterations. The file read and a single print happen once per run, so that overhead is tiny compared to the loop, and the results stay comparable in practice.
That said, I’m not against a compute-only / quiet mode. Since hyperfine already handles timing externally, the real work is implementing and maintaining a consistent --quiet / --no-io variant across 50+ languages.
If someone wants to contribute that (even starting with a subset), I’m happy to review PRs.
I'm not sure why the contents of rounds.txt isn't just provided as some kind of argument instead of read in from a file. Given all the other boilerplate involved, I would have expected it to be trivial to add relevant templating.
I'm not too familiar with the JVM so perhaps I'm missing something here: how would that help? The file is tiny, just a few bytes, so I'd expect the main slowdown to come from system call overhead. With non-mmap file I/O you've got the open/read/close trio, and only one read(2) should be needed, so that's three expensive trips into kernel space. With mmap, you've got open/stat/mmap/munmap/close.
Memory-mapped I/O can be great in some circumstances, but a one-time read of a small file is one of the canonical examples for when it isn't worth the hassle and setup/teardown overhead.
- There's a stark jump of times going from ~200ms to ~900ms. (Rust v1.92.0 being an in-between outlier.)
- C# gets massive boost (990->225ms) when using SIMD.
- But C++ somehow gets slower when using SIMD.
- Zig very fast*!
- Rust got big boost (630ms->230ms) upgrading v1.92.0->1.94.0.
- Nim (that compiles to C then native via GCC) somehow faster than GCC-compiled C.
- Julia keeps proving high-level languages can be fast too**.
- Swift gets faster when using SIMD but loses much accuracy.
- Go fastest language with own compiler (ie not dependent to GCC/LLVM).
- V (also compiles to C) expected it (appearing similar) be close to Nim.
- Odin (LLVM) & Ada (GCC) surprisingly slow. (Was expecting them to be close to Zig/Fortran.)
- Crystal slowest LLVM-based language.
- Pure CPython unsurpassable turtle.
Curious how D's reference compiler (DMD) compares to the LLVM/GCC front-ends, how LFortran to gfortran, and QBE to GCC/LLVM. Also would like to see Scala Native (Scala currently being inside the 900~1000ms bunch).
* Note that uses `@setFloatMode(.Optimized)` which according to docs is equivalent to `--fast-math` but only D/Fortran use this flag (C/C++ do not).
** Uses `@fastmath` AND `@simd`. The comparison supposedly is for performance on idiomatic code and for Julia SIMD is a simple annotation applied to the loop (and Julia may even auto do it) but should still be noted because (as seen in C# example) it can be big.
Author here! Thanks for the detailed breakdown. Let me address a few points:
- C++ SIMD being slower: The standard C++ uses i & 0x1 which lets the compiler auto-vectorize. With -O3 -ffast-math -march=native, gcc/clang do this really well. The explicit AVX2 version has overhead from manual vector setup and horizontal sum at the end. Modern compilers often beat hand-written SIMD for simple loops like this.
- Zig fast-math: Correct. Line 5 has @setFloatMode(.optimized) with a comment saying "like C -ffast-math".
- Julia: Also correct. Uses @fastmath @simd for - both annotations together.
- Crystal/Odin/Ada being slow: All three use x = -x which creates a loop-carried dependency that blocks auto-vectorization. The fast implementations use the branchless i & 0x1 trick instead.
- C# SIMD: Uses Vector512 doing 8 doubles per iteration. That explains the ~4x speedup.
- Nim vs C: Both compile via gcc with similar flags. Probably just measurement variance.
- Fortran: Interestingly does NOT use -ffast-math. Uses manual loop unrolling instead (processes 4 terms per iteration).
- Go: You're right that it's the fastest with its own compiler. No LLVM/GCC backend, just Go's own SSA-based compiler.
For suggestions - DMD, LFortran, and Scala Native would be great additions. PRs welcome!
Looking closer at the benchmarks, it seems that C# benchmark is not using AOT, so Go and even Java GraalVM get here an unfair advantage (when looking at the non SIMD versions). There is a non trivial startup time for JIT.
Sorry, I can't seem to edit my answer anymore, but I was mistaken, C# version is using AOT. But the are other significant differences here:
> var rounds = int.Parse(File.ReadAllText("rounds.txt"));
> var pi = 1.0D;
> var x = 1.0D;
> for (var i = 2; i < rounds + 2; i++) {
> x = -x;
> pi += x / (2 \* i - 1);
> }
> pi \*= 4;
> Console.WriteLine(pi);
For example, if we change the type of 'rounds' variable here from int to double (like it is also in Go version), the code runs significantly faster on my machine.
Try that on ARM64 and the result will be the opposite :)
On M4 Max, Go takes 0.982s to run while C# (non-SIMD) and F# are ~0.51s. Changing it to be closer to Go makes the performance worse in a similar manner.
This sort of thing is pretty meaningless unless the code is all written by people who know how to get performance out of their languages (and they're allowed to do so). Did you use the right representation of the data? Did you use the right library? Did you use the right optimization options? Did you choose the fast compiler or the slow one? Did you know there was a faster or slower one? If you're using fancy stuff, did you use it right?
I did the same sort of thing with the Seive of Eratosthenes once, on a smaller scale. My Haskell and Python implementations varied by almost a factor of 4 (although you could argue that I changed the algorithm too much on the fastest Python one). OK, yes, all the Haskell ones were faster than the fastest Python one, and the C one was another 4 times faster than the fastest Haskell one... but they were still over the place.
It's true this is a microbenchmark and not super informative about "Big Problems" (because nothing is). But it absolutely shows up code generation and interpretation performance in an interesting way.
Note in particular the huge delta between rust 1.92 and nightly. I'm gonna guess that's down to the autovectorizer having a hole that the implementation slipped through, and they fixed it.
Take their Haskell implementation. If I compile it unoptimized (with a slight change to hardwire the iteration count at one billion instead of reading it from a file), I get bored waiting for it to finish after several minutes of clock time. If I compile it with "-O3", it runs in 4.75 seconds (it's an old machine).
Suppose I remove the strictness annotations (3 exclamation points, in places that aren't obvious to a naive programmer coming from almost any other language). If I then compile it unoptimized, it gets up to over 30GB of resident memory before I get bored (it's an old machine with a lot of memory). It would probably die with an out of memory error if I tried to run it to completion. However, if I compile that same modified code optimized, the compiler infers the strictness and the program runs in exactly the same time as it does with the annotations there. BUT it's far from obvious to the casual observer when the compiler can make those inferences and when it can't.
I had ChatGPT rewrite the Haskell code to use unboxed numeric types. It ran in 1.5 seconds (the C version takes 1.27). The rewrite mostly consists of sprinkling "#" liberally throughout the code, but also requires using a few specialized functions. I had ChatGPT do it because I have never used unboxed types, and you could argue that they're not common idiom. However, anybody who actually wrote that kind of numerical code in Haskell on a regular basis would use unboxed types as a matter of course.
You're acting like this is a gotcha, but the answer is obviously "all of them" and that indeed, this tells you interesting things about the behavior of your compiler. There are lots of variant scores in the linked article that reflect different ways of expressing the problem.
But also, it tells you something about the limitations of your language too. For example, the biggest single reason that C/C++ (and languages like Fortran/Zig/D and sometimes C# and Rust whose code generation is isomorphic to them) sit at the top of the list is that they autovectorize. SIMD code isn't a common idiom either, but the compiler figures it out anyway.
And apparently Haskell isn't capable of doing enough static analysis to fall back to an unboxed implementation (though given the simplicity of this code, that should be possible I think). That's not a "flaw" and it doesn't mean Haskell "loses", but it absolutely is an important thing to note. And charts like this show us where those holes lie.
> You're acting like this is a gotcha, but the answer is obviously "all of them" and that indeed, this tells you interesting things about the behavior of your compiler.
They tell me interesting things if I know enough about the language to know the difference. It tells me things if I'm getting into the weeds with Haskell specifically. That doesn't make the big comparison chart useful in any way.
I still don't know anything that lets me compare anything with any other language unless I actually know that language nearly as well. And I definitely don't get much out of a long list of languages, most of which I know not at all or at most at a "hello world" level, with only a couple of the entries tagged with even minimal information about compilers or their configurations at all. Especially when, on top of that, I don't know how much the person writing the test code knew.
At most I get "this language does a pretty good/poor job on this type of task when given code that may or may not be what a 'native expert' would write.".
And that's not news. Nobody (with any sophistication) would write that code for real in Python, or probably in Haskell either, because most seasoned programmers know that if you want speed on a task like that, you write it in a more traditional compiled procedural language. It's also not a kind of code that most people write to begin with. If you want an arctangent (which is really what it's doing), you use the library function, and the underlying implementation of that is either handcrafted C, or, more likely, a single, crafted CPU instruction with some call handling code wrapped around it.
So what is the overall chart giving me that I can use?
> So what is the overall chart giving me that I can use?
"If you write in C or an analog, your math will autovectorize nicely"
"If you use a runtime with a managed heap, you're likely to take a penalty even on math stuff that doesn't look heap limited"
"rust 1.92 is, surprisingly, well behind clang on autovectorizable code"
I mean, I think that stuff is interesting.
> If you want an arctangent (which is really what it's doing), you use the library function
If you just want to call library functions, you're 100% at the mercy of whatever platform you picked[1]. So sure, don't look at benchmarks, they can only prove you wrong.
[1] Picked without, I'll note, having looked carefully at cross-platform benchmarks before having made the choice! Because you inexplicably don't think they tell you anything useful.
The delta there is because the Rust 1.92 version uses the straightforward iterative code and the 1.94-nightly version explicitly uses std::simd vectorization. Compare the source code:
> Note in particular the huge delta between rust 1.92 and nightly. I'm gonna guess that's down to the autovectorizer having a hole that the implementation slipped through, and they fixed it.
The benchmark also includes startup time, file I/O, and console printing. There could have been a one-time startup cost somewhere that got removed.
The benchmark is not really testing the Leibniz loop performance for the very fast languages, it's testing startup, I/O, console printing, etc.
Sure, but at the end of the day, youre still just polishing a turd, and you give up a LOT of ecosystem just to get a still deeply unimpressive benchmark time.
After seeing Swift's result I had to look into the source to confirm that yes, it was not written by someone who knows the language.
But this is a good benchmark results that demonstrate what performance level can you expect from every language when someone not versed in it does the code porting. Fair play
Author here. There are actually 3 Swift versions in the benchmark:
- Swift (standard): 893ms
- Swift (relaxed): 903ms (uses fast-math equivalent)
- Swift (SIMD): 509ms (explicit SIMD4)
The standard version uses x *= -1.0 which creates a loop-carried dependency that blocks auto-vectorization - same issue as Crystal, Odin, Ada. The SIMD version uses the branchless i & 0x1 trick and is ~1.75x faster.
Fair point that someone versed in Swift would probably use the better pattern in the standard version too. PRs welcome to improve it! The goal was idiomatic-ish code, but I'm not an expert in all 40+ languages.
and I would hope/expect the compiler to be smart enough to know that it only has to check “rounds+2” once there. Swift isn’t exactly new anymore, and it’s supported by a large company.
Startup time doesn’t seem to be factored in correctly, so any language that uses a bytecode (e.g. Java) or is compiling from source (e.g. Ruby, Python, etc.) will look poor on this. If the kids of applications that you write are ones that exit after a fraction of a second, then sure, this will tell you something. But if you’re writing server apps that run for days/weeks/months, then this is useless.
Python took 86 seconds, if I'm reading it correctly. I can see your point holding for a language like Java, but most of Python's time spent cannot have been startup time, but actual execution time.
Reading the fine print, the benchmark is not just the Leibniz formula like it says in the chart title. It also includes file I/O, startup time, and console printing:
> Why do you also count reading a file and printing the output?
> Because I think this is a more realistic scenario to compare speeds.
Which is fine, but should be noted more prominently. The startup time and console printing obviously aren't relevant for something like the Python run, but at the top of the chart where runs are a fraction of a second it probably accounts for a lot of the differences.
Running the inner loop 100 times over would have made the other effects negligible. As written, trying to measure millisecond differences between entire programs isn't really useful unless someone has a highly specific use case where they're re-running a program for fractions of a second instead of using a long-running process.
The Clojure version is not AOT'd, so it's measuring startup + compiler time. When properly compiled it should be comparable to the Java implementation.
I’m genuinely blown away by all the interest in what started as a silly little experiment.
The project grew way beyond its original scope.
My initial curiosity was simply: how could you set up a pipeline to do automatic speed comparisons? I was less interested in the results as a definitive measure and more in the infrastructure challenge itself.
But then the interest kept growing. I tried to modernize things, but one thing became quite notable: the difference between a language that gets actively optimized by its community (like Julia) versus one that just sits there unoptimized is striking.
Honestly, I got overwhelmed. Managing all those implementations, keeping versions up to date, reviewing contributions—it was a lot. I basically tapped out for about a year.
Now I’m back, and with AI assistance, maintaining this has become much more realistic—updating versions, helping optimize implementations, etc. That said, I’m always happy to accept contributions from folks who know their languages better than I do.
Thank you all for your interest and the thoughtful discussion!
To explain what I meant. I knew that that in the end this was a microbenchmark. It can certainly give you some clue about a language, but it doesn't tell you the whole picture. In the end it tells you how good a language is (or can be) at loops and floating point math.
That's what I meant. I hope that makes it clearer.
Thanks for the links - the benchmarks game is a great resource and has been around for a long time.
That said, I'm not sure what "do better" means here. This is an open source project I maintain in my spare time. If you see room for improvement, PRs are always welcome. That's how open source works - if something is useful to you and you want it improved, contribute.
The project exists because some people find it useful. If it's not for you, that's fine too.
This is meaningless. The benchmarks are (1) run in github actions, (2) include file and console IO, and (3) are compiled with different compiler flags...
It's certainly not meaningful of anything in the first place, since it only tests codegen for a sequence of floating point calculations, which is hardly a representative workload for anything. Even if it were, using an unreliable virtualised environment only introduces noise, especially in the presence of syscalls... So it's worthless even if the entire purpose was to evaluate how fast floating point sums and divisions run in a CI server x)
C# non SIMD (naive non optimized version) is in the same ballbark as other similar GC languages. Nim version is not some naive version also and seem rather specially crafted so it can be vectorized and still looses to C# SIMD.
Loses? My comparison is regarding GP's metric perf/lines_of_code. Let m := perf/lines_of_code = 1/(t × lines_of_code) [highest is better], or to make comparison simpler*, m' := 1/m = t × lines_of_code [lowest is better]. Then**:
Nim 1672
Julia 3012
D 3479
C# (SIMD) 5853
C# 8919
>Nim version is not some naive version
It's direct translation of formula, using `mod` rather `x = -x`.
*Rather comparing numbers << 1.
**No blank/comment lines. As cloc and similar tools count.
Nim "cheats" in a similar way C and C++ submissions do: -fno-signed-zeros -fno-trapping-math
Although arguably these flags are more reasonable than allowing the use of -march=native.
Also consider the inherent advantage popular languages have: you don't need to break out to a completely niche language, while achieving high performance. Saying this, this microbenchmark is naive and does not showcase realistic bottlenecks applications would face like how well-optimized standard library and popular frameworks are, whether the compiler deals with complexity and abstractions well, whether there are issues with multi-threaded scaling, etc etc. You can tell this by performance of dynamically typed languages - since all data is defined in scope of a single function, the compiler needs to do very little work and can hide the true cost of using something like Lua (LuaJIT).
It would be nice to have a different color for languages that give the exact IEEE 754 result. Other languages can be using some kind of fast math. Sorted by accuracy:
# Language Accuracy
14 Swift (SIMD) 8.69
9 Fortran 90 9.44
2 C# (SIMD) 9.49
. [All the others] 9.50
// [I removed the error checking]
def pi_accuracy(value_str):
"""Calculate accuracy of computed pi value.
Returns the number of correct decimal places (higher is better).
"""
value = float(value_str)
# math.pi is available in MicroPython
accuracy = 1 - (value / math.pi)
return -math.log10(abs(accuracy))
Some implementations seem vectorization-friendly like the C one that uses a bit-twiddling trick to avoid the `x = -x` line that the Odin implementation and others have.
When you put these programs into Godbolt to see what's going on with them, so much of the code is just the I/O part that it's annoying to analyze
I think I get why C++ thru C are all similar (all compile to similar assembly?), but I don't get why Go thru maybe Racket are all in what looks like a pretty narrow clump. Is there a common element there?
The common element is that they're written with the most obvious version of the code, while the ones in the faster bucket are either explicitly vectorized or written in non-obvious ways to help the compiler auto-vectorize. For example, consider the Objective C version of the loop in leibniz.m:
for (long i = 2; i <= rounds + 2; i++) {
x *= -1.0;
pi += x / (2.0 * i - 1.0);
}
With my older version of Clang, the resulting assembly at -O3 isn't vectorized. Now look at the C version in leibniz.c:
rounds += 2u; // do this outside the loop
for (unsigned i=2u; i < rounds; ++i) // use ++i instead of i++
{
double x = -1.0 + 2.0 * (i & 0x1); // allows vectorization
pi += (x / (2u * i - 1u)); // double / unsigned = double
}
This produces vectorized code when I compile it. When I replace the Objective C loop with that code, the compiler also produces vectorized code.
You see something similar in the other kings-of-speed languages. Zig? It's the C code ported directly to a different syntax. D? Exact same. Fortran 90? Slightly different, but still obviously written with compiler vectorization in mind.
(For what it's worth, the trunk version of Clang is able to auto-vectorize either version of the loop without help.)
I think it's SIMD generation. Managed runtimes have a much harder time autovectorizing, because you can't do any static analysis about things like array sizes. Note that the true low-level tools are all clustered around 2-300ms, and that the next level up are all the "managed" runtimes around 1-2s.
The one exception is sort of an exception that proves the rule: it's marked "C# (SIMD)", and looks like a native compiler and not a managed one.
They’re measuring program execution time, including program startup and tear down. Languages with a more complex runtime take longer for the startup, and all seem to have roughly equally optimized that.
That first big jump in the graph, I thought that it must be the divide between auto-gc'd and non auto-gc'd languages. But then I noticed that Rust is on the wrong side of the divide.
The code looks 100% identical except for the namespace prefixes. Must be something particular about github setup, because on mine (gcc15.2.1/clang20.1.8/Ryzen5600X) the run time is indistinguishably close. Interestingly, with default flags but -O3 clang is 30% slower, with flags from the script (-s -static -flto $MARCH_FLAG -mtune=native -fomit-frame-pointer -fno-signed-zeros -fno-trapping-math -fassociative-math) clang is a bit faster.
A nitpick is that benchmarking C/C++ with $MARCH_FLAG -mtune=native and math magic is kinda unfair for Zig/Julia (Nim seem to support those) - unless you are running Gentoo it's unlikely to be used for real applications.
I suspect this is it. Any benchmark that takes less than a second to run should have its iteration count increased such that it takes at least a second, and preferably 5+ seconds, to run. Otherwise CPU scheduling, network processing, etc. is perturbing everything.
BenchExec "uses the cgroups feature of the Linux kernel to correctly handle groups of processes and uses Linux user namespaces to create a container that restricts interference of [each program] with the benchmarking host."
Certainly better, but you’re always going to be better off maximizing the runtime to a level where it just swamps any of the other effects. Then do multiple runs and take an average.
(I have spent a good amount of time hacking the llvm pass pipeline for my personal project so if there was a significant difference I probably would have seen it by now)
You are correct, that was an uneducated guess on my part.
I just glanced at the IR which was different for some attributes (nounwind vs mustprogress norecurse), but the resulting assembly is 100% identical for every optimization level.
As with all the ahead-of-time compiled languages that I checked, the answer is that it generates non-SIMD code for the hot loop. The assembly code I see in godbolt.org isn't bad at all; the compiler just didn't do anything super clever.
In my opinion, the comparisons could be better if the file I/O and console printing were removed.
reply