Each Benchmarks

Elementwise sum and scale bandwidth benchmarks comparing NumKong against scalar baselines, ndarray, and nalgebra.

Rust

Library	Precision	GB/s
Sum
numkong::EachSum	f32 → f32	97.55
nalgebra::add	f32 → f32	95.31
ndarray::add	f32 → f32	94.84
serial code	f32 → f32	94.06
serial code	f64 → f64	85.48
ndarray::add	f64 → f64	84.91
nalgebra::add	f64 → f64	84.55
numkong::EachSum	f64 → f64	82.77
numkong::EachSum	f16 → f16	96.56
numkong::EachSum	bf16 → bf16	17.73
numkong::EachSum	i8 → i8	111.47
serial code	i8 → i8	110.81
Scale
serial code	f32 → f32	82.22
ndarray::scale	f32 → f32	81.75
numkong::EachScale	f32 → f32	66.56
nalgebra::scale	f32 → f32	39.52
serial code	f64 → f64	72.46
ndarray::scale	f64 → f64	72.39
numkong::EachScale	f64 → f64	66.70
nalgebra::scale	f64 → f64	38.58
numkong::EachScale	f16 → f16	66.23
numkong::EachScale	bf16 → bf16	33.19
serial code	i8 → i8	89.21
numkong::EachScale	i8 → i8	26.43

Python

Library	Precision	GB/s
Sum
numpy.add	i8 → i8	143.56
numkong.add	i8 → i8	123.77
numkong.add	f32 → f32	118.39
numpy.add	f32 → f32	115.32
numpy.add	f64 → f64	114.37
numkong.add	f16 → f16	107.29
numkong.add	f64 → f64	100.01
numkong.add	bf16 → bf16	73.27
numpy.add	f16 → f16	4.08

Run It

Rust

# Default 1M-element tensors
cargo bench --bench bench_each --features bench_each

# Focus on one operation family
NUMWARS_FILTER="each/sum|each/scale" \
cargo bench --bench bench_each --features bench_each

Python

# Default 1M-element tensors, add on float32
python each/bench.py --filter 'add/float32'