zephyr — benchmarks

ZEPHYR v0.2 — zc terminal

cat benchmarks.txt

======================================================
 BENCHMARKS — faster than C, checksums shown
 best-of-7 · zen5 · windows x86-64
======================================================

The same algorithm runs in Zephyr, C (gcc -O2), and Rust (rustc -O). All three must produce a byte-identical result before a single timing counts.

Zephyr wins where it hunts for loops LLVM leaves on the table — matmul's inner loop gets a hand-written 4-wide AVX2 kernel where LLVM settles for 2-wide SSE2. Not universally faster. Verifiably faster here, and honest about the rest.

bench --matmul --cube

Wins outright.

fig.1 — integer matmul, 768×768

Zephyr42 ms

Rust70 ms

C72 ms

fig.2 — wireframe cube, 12k frames

Zephyr8 ms

C10 ms

Rust11 ms

bench --all --best-of 7

Every workload.

Lower is faster. Zephyr's column is highlighted; the verdict names the closest competitor.

Workload	Zephyr	C −O2	Rust −O	Result
integer SIMD (matmul)	42	72	70	wins both
rasterization (cube)	8	10	11	wins both
bignum (pi)	66	84	91	wins both
alloc / GC (strings)	132	172	148	wins both
sorting	139	276	43	2× faster than C
recursion (fib)	21	12	21	ties Rust
hash map	83	32	69	ties Rust
float compute (mandel)	109	88	90	within 1.2×

all times in ms · peak memory lowest of the three on most rows · compiles ~10× faster than gcc/rustc

./bench/run_all.ps1

Rerun it yourself.

$ .\bench\run_matmul.ps1
$ .\bench\run_cube.ps1
  # checksum verified, then timed, best-of-7

toolchain docs back to home