Adding A Binarized CNN Accelerator To RISC V For Person Detection
summary
A person detector runs on a tiny FPGA board with ~1 Mbit SRAM and a VGA camera downscaled to 32×32 pixels, using an Orca RISC‑V core plus binary CNN accelerator and streaming vector extensions. The complete system is under ~5,000 LUTs, draws ~5 mW, performs ~80 convolution ops per cycle at 24 MHz and reports ~71× speedup versus a scalar core. An int8 model (40× fewer ops) matches float32 accuracy on the person class.
AI-written summary · about
chapters (15)
- 0:00Intro
- 1:08Smarter Sleep
- 2:19Smarter Doors
- 2:42Person Detector Demo
- 3:21Deep Learning Database People
- 3:50Person Detector System
- 4:42Streaming Vectors
- 5:57Streaming Vector Extensions
- 6:56Extended SVE
- 7:41FIR filter (12x speedup)
- 8:22Custom Streaming Accelerators
- 8:58Streaming Convolution Accelerator
- 9:41Person Detector Performance
- 10:10Area Breakdown
- 11:23SVE vs "RISC-V Vector Extension"
related talks
11mRISC-V enabled, low-power CNN classification in Edge devices
12mCustomizing RISC-V Cores to Accelerate Neural Networks
30mAccelerating Neural Networks for Vision Systems via FPGAs
13mIMU Sensor ML Action Detection Optimization with Custom Instructions on Embedded RISC-V Cores
01h 11mMachine Learning Acceleration in an Embedded World
07mDemo: Neural Network Acceleration on Metis, Powered by RISC-V
17mNext-Generation Edge AI with RISC-V Vector Cores for Vision Applications
13mKeynote: Taking RISC-V Intelligence Everywhere