# Maratyszcza/NNPACK

Acceleration package for neural networks on multi-core CPUs

Repository: https://github.com/Maratyszcza/NNPACK
Canonical: https://ross.abutalabs.com/products/nnpack
Language: C
License: BSD-2-Clause
License Family: permissive
Topics: neural-network, neural-networks, convolutional-layers, inference, high-performance, high-performance-computing, simd, cpu, multithreading, fast-fourier-transform, winograd-transform, matrix-multiplication
Last push: 2024-06-11T14:51:27+00:00

## Health v2 (maintenance only)
Score: 32/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 0, release rhythm 35, longevity 100
- inputs: {"age_days": 3817, "days_push": 813, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1710, forks 321 (observed 2026-08-28T04:05:25.611861+00:00)

## What it is
NNPACK is a C99 acceleration package providing high-performance SIMD and multi-core CPU implementations of neural network layers, especially convolutional layers, using FFT, Winograd, and implicit GEMM algorithms. It serves as a low-level performance primitive library consumed by deep learning frameworks like PyTorch, Caffe2, and MXNet rather than being used directly by researchers.

## Use cases
- speed up convolutional layer inference on multi-core CPUs
- accelerate convnet forward and backward propagation with SIMD
- use FFT-based convolution for large kernels
- use Winograd transform for fast 3x3 convolutions
- integrate optimized NN primitives into a deep learning framework
- run neural network inference on ARM mobile devices with NEON

## When to choose
- you need maximum CPU performance for convnet layers in a framework or runtime
- you target x86-64 with AVX2 or ARM with NEON, including mobile and WebAssembly
- you want dependency-free C99 primitives for convolution and fully-connected layers

## When to avoid
- you want a high-level ML framework or training toolkit rather than low-level primitives
- you need GPU acceleration
- you need actively developed features or modern CPU instruction sets beyond AVX2, as the project is in maintenance mode

## Facets
- artifact type: library
- maturity: maintenance
- function: machine-learning, llm-inference, benchmarking, concurrency
- domain: deep-learning, machine-learning, performance, gpu-computing
- platform: wasm, cpp, python
- tags: simd, neural-networks, convolution, cpu-optimization, winograd-transform, fft, inference-primitives, c99, linux, macos, android, ios

## Member repositories
- Maratyszcza/NNPACK (main) score 32

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:25.611861+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:35:02.807139+00:00, confidence not recorded.
  - readme: https://github.com/Maratyszcza/NNPACK (fetched 2026-08-28T04:05:25.611861+00:00, sha 2795bf396439)
- Data as of 2026-08-30T08:39:29.467469+00:00.