April 6, 2026
Benchmarks, bluster, big feelings
Idiomatic Koru Kernels Match Hand-Specialized C
Koru says it’s as fast as C; the internet yells: prove it
TLDR: Koru claims its code runs as fast as hand‑tuned C, and benchmarks show a tie with a highly optimized C version. Commenters are split between hype, “prove it” skeptics demanding machine‑code receipts, and jokesters roasting the project’s grandiose vibe — turning a speed test into a full‑blown internet brawl.
A new post claims Koru’s “kernels” — short bits of number‑crunching code — can run as fast as hand‑tuned C, the gold standard for speed. In quick tests, plain Koru beat ordinary C, Rust, and Zig versions, then tied a super‑optimized C rewrite. The pitch: Koru carries enough meaning that its compiler makes smart choices automatically, no magic flags required. All the code is on GitHub — but the comment section immediately turned into a spectator sport.
The crowd split fast. Some were howling at the project’s over‑the‑top tagline, with one user asking if the whole thing was a parody. Others went scorched earth: one called the post “slop” and joked that if AI wrote the blog, why not let it do the optimization, too. A more measured camp shrugged that this looks like what a good compiler does anyway — like “autovectorization,” a computer trick that automatically speeds loops — while purists demanded the ultimate receipts: show the generated machine code or it didn’t happen. Then came the twist: a commenter surfaced the author’s own confession that this entire saga began as “shitposting” about Lisp being slow, which summoned a rival to prove otherwise. Cue the popcorn.
So yes, Koru might run fast. But the real race is in the comments, where memes about “postmodern zero‑cost fractals” are neck‑and‑neck with “prove it” skeptics. Tech drama, start your engines.
Key Points
- •Koru’s fused n-body kernel matched the performance of expert hand-specialized C on a five-body problem.
- •Initial comparisons showed Koru (1.00) outperforming naive idiomatic C (1.14), Rust (1.17), and Zig (1.14) under identical benchmarking conditions.
- •A follow-up fixed-size scalarized C implementation, compiled with and without -ffast-math, achieved 1.00 and matched Koru.
- •The result is attributed to Koru kernels carrying semantic information that enables compilers to generate highly optimized code without manual specialization.
- •Benchmarks used 50M iterations, warmup=3, multiple runs via hyperfine, same hardware/harness; code is available on GitHub, and numeric ratios are approximate but ordering is stable.