CMSIS-NN
summary
CMSIS-NN provides optimized kernels and a toolchain (the Arm NN SDK) to run neural nets on Cortex‑M microcontrollers by quantizing weights to 8-bit and relayouting data. A CRNN example fits ~80 KB of parameters and uses ~3 million ops; a depthwise-separable variant halves memory but roughly doubles ops. Using CMSIS-NN operators yields about 5× speed and energy improvements. Demo runs on an OpenMV Cortex‑M7 board classifying images at ~4 fps (monochrome ~12 fps).
AI-written summary · about
chapters (15)
- 0:00Introduction to CMSIS-NN
- 1:07Why run ML at the edge
- 4:37ML on Cortex-M microcontrollers
- 6:48Deploying neural networks on Cortex-M
- 9:16Choosing the right neural net model
- 12:06Compressing models for deployment
- 13:54ARM NN SDK and tools
- 14:59Optimizing with CMSIS-NN
- 16:50Live demo of animal recognition
- 20:37Q&A on model compilation
- 21:44Q&A on cross-platform support
- 22:06Q&A on keyword recognition
- 22:51Q&A on preprocessing
- 23:55Q&A on object detection
- 24:17Closing remarks
related talks
24mRunning AI/Neural networks on microcontrollers made simple with the STM32Cube.AI
19mOptimizing CNN Model Inference on CPUs
28mML on the Edge for Industry 4.0 with Arm Ethos N-78 Neural Processing Unit
07mDemo: Neural Network Acceleration on Metis, Powered by RISC-V
38mNEURAL NETWORKS ON YOUR MOBILE PHONE. DIVING INTO THE NN API
01h 02mMulti-Dimension Mapping of Embedded Applications
26mMemory-Efficient LLM Inference on Edge Devices With NNTrainer
12mCustomizing RISC-V Cores to Accelerate Neural Networks