Customizing RISC-V Cores to Accelerate Neural Networks

on YouTube ↗12m

" summary

Demonstration shows the Codasip L31 core extended with a small AI accelerator and custom instructions (notably a pushconf instruction plus a FIFO register file) that speed MNIST convolutional kernels, implemented alongside TensorFlow Lite Micro. Code profiling in Codasip Studio drives selection, the custom convolution fits in about 200 lines of code, and the result claims ~5x runtime and ~3x lower power in the demo (uploader summary states >3x speed and halving energy). Tools include Codal HDL and Codasip Studio for simulation and estimation.

AI-written summary · about

" chapters (10)

1. 0:00Introduction to Codasip
2. 0:36Company background
3. 0:59Benefits of edge AI
4. 1:43Challenges in edge processing
5. 2:30Profiling and optimization
6. 3:44Architecting the solution
7. 5:50Codal language and modeling
8. 8:00Implementation details
9. 9:28Performance results
10. 10:49Conclusion and demo info

" related talks

• 13mDemo: RISC-V as an Enabler of Heterogeneous ComputeZdenek Prikryl · RISC-V Summit North America 2023
• 10mAI custom Software/Hardware Interface improving performance 5-10xKeith Graham · RISC-V Summit Europe 2024
• 01h 10mWhy RISC-V Through ISA Customization DemonstrationKeith Graham · RISC-V Summit North America 2023
• 05mDemo: Optimizing ML for IMU Sensor Action DetectionPeter Robertson, Thomas Hepworth · RISC-V Summit North America 2023
• 11mA RISC V ISA Extension for Ultra Low Power IoT Wireless Signal ProcessingZdenek Prikryl · RISC-V Summit 2019
• 20mBoosting AI on Semidynamics RISC-V Cores with Custom Tensor Instr...Roger Espasa, José María Arnau · RISC-V Summit North America 2023
• 07mDemo: Neural Network Acceleration on Metis, Powered by RISC-VFlorian Zaruba, Victor Labian Carro · RISC-V Summit North America 2025
• 09mA Tightly Coupled Light Weight Neural Network Processing Units With RISC V CoreLei Zhang · 7th RISC-V Workshop 2017