NumKong: 2'000 Mixed Precision Kernels For All
Summary
NumKong is a massive open-source release featuring over 2,000 SIMD kernels for mixed-precision numerics, spanning seven languages and major ISA families. The post dives into kernel design patterns, performance benchmarks, and practical workflows (including geospatial, alignment, and late-interaction scoring) to illustrate how tiled GEMMs and fused epilogues can accelerate AI workloads on CPUs and WASM runtimes.