IMU Sensor ML Action Detection Optimization with Custom Instructions on Embedded RISC-V Cores

on YouTube ↗13m

" summary

Small IMU-based action detector (standing/walking/running) is a quantized RNN with under 6,000 parameters running on a tiny single-stage RISC-V core customized in Codasip Studio. Custom instructions implemented: MAC32, load-with-post-increment, a saturation/shift fusion, and two-way 16-bit SIMD MAC+accumulate. Results: per-inference cycles drop ~35% overall (13% then 23% from fused ops), ~1.6× inference speed from SIMD, ~30% power reduction, ~11% area increase. Design is validated on FPGA and synthesized for TSMC 28nm.

AI-written summary · about

" chapters (5)

1. 0:00Project introduction
2. 2:06Workload and model design
3. 4:04Toolchain and methodology
4. 6:30Implementation and optimization
5. 11:26Results and conclusion

" related talks

• 10mAI custom Software/Hardware Interface improving performance 5-10xKeith Graham · RISC-V Summit Europe 2024
• 14mLow-Cost SIMD Module for ML AccelerationLeonidas Kosmidis, Marc Solé Bonet · RISC-V Forum: Vector & Machine Learning 2021
• 11mGetting Started with RISC-V Custom InstructionsJon Taylor · embedded world 2023
• 12mCustomizing RISC-V Cores to Accelerate Neural NetworksJon Taylor · embedded world 2022
• 40mHands-on TinyML Deployment on RISC-V: Building an Optimized Edge AI Pipeline - Corrected AudioYueh-Feng Lee, Niraj Dengale · RISC-V Summit Europe 2026 Developer Workshops
• 13mAdding A Binarized CNN Accelerator To RISC V For Person DetectionGuy Lemieux · 7th RISC-V Workshop 2017
• 10mBoosting AI on Semidynamics RISC-V Cores with Custom Tensor InstructionsRoger Espasa · RISC-V at embedded world 2024
• 09mDeveloping Custom RISC-V ISA Extensions for General Embedded Image Processing OperationsCarsten Rolfes, Stephan Nolting · RISC-V Summit Europe 2025