Five megabytes of math that widens before it overflows

What it does

NumKong is a portable numeric kernel library that provides SIMD-accelerated dot products, distance metrics, matrix multiplication, geospatial functions, and cast/reduce operations across 15+ numeric types—from 4-bit integers and 6-bit floats to 128-bit complex—on x86, Arm, RISC-V, LoongArch, Power, and WebAssembly. It exposes these through bindings for C, C++, Python, Rust, JavaScript, Swift, and Go, and deliberately avoids managing threads or memory so it slots into whatever parallelism model or allocator you already use. Every kernel is validated against 118-bit extended-precision baselines and cross-checked against OpenBLAS, Intel MKL, and Apple Accelerate.

The interesting bit

Where most libraries return dot products in the same type as the inputs—Float16 × Float16 → Float16, Int8 × Int8 → Int8—NumKong widens accumulators to dodge quiet overflow. That gives it a latency edge over NumPy, PyTorch, and JAX on single-threaded dot products, though throughput on large f64 or f32 matrix multiplies still favors mature BLAS. The payoff is binary size: roughly 5 MB versus PyTorch’s 705 MB or JAX’s 357 MB.

Key highlights

• 2,000+ SIMD kernels covering 1D and 2D registers like AMX and SME.
• 15+ numeric types with hardware-aware defaults (Arm prefers f16, x86 prefers bf16).
• Wider accumulators by default: i8 → i32, f16 → f32, f32 → f64, plus compensated Dot2 summation for f64.
• Higher single-threaded latency and lower error than NumPy/OpenBLAS, PyTorch/MKL, and JAX on many types in the provided benchmarks.
• 5 MB binary with no hidden allocations or thread management.

Caveats

• Throughput-oriented matrix multiplication is not uniformly faster than mature BLAS; on the tested Intel Sapphire Rapids chip, NumKong trails PyTorch + MKL for f64 and f32 matrix multiply, though it leads for i8 and e5m2.
• Language bindings are uneven: JavaScript lacks geospatial, mesh, sparse, curved, reduce, each, and trigonometry ops; Swift and Go are also missing several categories.

Verdict

NumKong is worth a look if you need low-latency, mixed-precision numeric kernels in a small, embeddable binary without hidden memory or thread behavior. If you are already saturating PyTorch or MKL on throughput-heavy f64/f32 matrix work and don’t mind the install size, it may not replace your current stack.