Demo: Optimizing ML for IMU Sensor Action Detection
Talkat RISC-V Summit North America 2023 · on Tuesday, 2023-11-07· at 13:05· Expo Hall - Exhibit Hall A - Demo Theater·
summary
Codasip Studio customizes a reference Single Stage Core to run a quantized neural network for IMU activity detection, adding concrete custom instructions like MAC32 and load-post-increment that the SDK applies automatically. Two new instructions cut cycles per inference by about 13% and reduce power slightly; the full set of custom instructions produces roughly a 1.6x speedup and ~30% lower inference power with only a small area increase. The compiler emits and uses the new ops in real code.
AI-written summary · about
related talks
12mCustomizing RISC-V Cores to Accelerate Neural Networks
45mUsing Edge Machine Learning to save power and development time
25mCMSIS-NN
07mDemo: Neural Network Acceleration on Metis, Powered by RISC-V
24mRunning AI/Neural networks on microcontrollers made simple with the STM32Cube.AI
26mNext generation machine learning for mobile and embedded platforms
46mHackathon Presentations
14mLow-Cost SIMD Module for ML Acceleration