“Jumpstart Your Edge AI Vision Application with New Development Kits from Avnet,” a Presentation from Avnet
0 likes41 views
Avnet introduces new development kits for edge AI vision applications, featuring the RZBoard V2L and I.MX 8M Plus Edge AI Kit, designed for energy-efficient machine vision. These kits support various applications such as object detection, stereo vision, and facial recognition while utilizing advanced processors and camera modules for enhanced performance. Additionally, the Zuboard 1CG offers a low-cost solution for machine vision with versatility in hardware configurations and expansive model support.
![Renesas RZ/V2L Power Saving Features
1. Detect zeros to prevent
unnecessary operations
2. Reuse data for matrix
operations
7
0
5
10
15
20
25
30
Resnet-50 MobileNet v1 MobileNet v2 TinyYOLOv2 TinyYOLOv3 YOLOv3
Power Efficiency[fps/W]
RZ/V2M
RZ/V2L
A社
B社
Power Efficiency (fps/W)
© 2022 Avnet](https://image.slidesharecdn.com/e1w05houstonavnet2022-220627131349-33dbfb4c/95/Jumpstart-Your-Edge-AI-Vision-Application-with-New-Development-Kits-from-Avnet-a-Presentation-from-Avnet-7-638.jpg)
“Jumpstart Your Edge AI Vision Application with New Development Kits from Avnet,” a Presentation from Avnet
- 1. Jumpstart Your Edge AI Vision Application with New Development Kits from Avnet Monica Houston Manager, Technical Solutions Engineering Avnet
- 2. RzBoard V2L Machine Vision Avnet SBC Boards 2 © 2022 Avnet 8M Plus Edge AI Kit ZUboard 1CG
- 3. Avnet IAS Camera Adapters 3 Dual IAS camera adapter • 2x onsemi AR1335 IAS camera modules • 1x onsemi AP1302 ISP coprocessor • Supports stereo 4K video (4K 30 fps, 1080P 60 fps, 720P 120 fps) SYZYGY dual IAS camera adapter • 2x onsemi AR0144 IAS camera modules • 1x onsemi AP1302 ISP coprocessor Pi IAS camera adapter • 1x onsemi AS0260 IAS camera module (includes onchip ISP coprocessor) • Supports HD video (1080p / 30 fps or 720p / 60 fps) © 2022 Avnet
- 4. Energy-efficient Vision AI Using RzBoard V2L
- 5. What is RzBoard V2L? • Avnet-designed development board, engineered in compact Raspberry Pi-4 (MaaXBoard) form-factor • Facilitates development of Vision-AI applications based on energy-efficient Renesas RZ/V2L SoC • Efficient on-chip DRP-AI accelerator (1 TOPS/W), capable of Tiny YOLOv2 object-detection at 28 fps • Supports Mikroe click boards and Raspberry Pi Hats 5 Applications • Barcode scanning • Object detection • Pose detection • SLAM © 2022 Avnet
- 6. Renesas RZ/V2L DRP-AI • DRP (Dynamically Reconfigurable Processor): Programable hardware • AI-MAC (multiply-and-accumulate): Hardware dedicated for MAC computing • DMAC (Direct Memory Access Controller) DRP-AI Translator: • Tool to generate DRP-AI optimized executables from trained ONNX model • Optimize the graph structure Develop and run applications on Yocto Linux 6 © 2022 Avnet
- 7. Renesas RZ/V2L Power Saving Features 1. Detect zeros to prevent unnecessary operations 2. Reuse data for matrix operations 7 0 5 10 15 20 25 30 Resnet-50 MobileNet v1 MobileNet v2 TinyYOLOv2 TinyYOLOv3 YOLOv3 Power Efficiency[fps/W] RZ/V2M RZ/V2L A社 B社 Power Efficiency (fps/W) © 2022 Avnet
- 8. Smart Retail — Object Detection Demo 8 RzBoard V2L Dev Board RZ/V2L #Core1 Linux #Core0 HDMI MIPI-CSI ISP Image for DRP-AI DRP-AI MIPI-CSI Camera Monitor HDMI AI Inference Result (Object name, and coordinate) Display Picture Size：540P* Input Image at 30 fps captured resolution: 1080P Frame Rate: 30 fps output resolution: 540P* Frame Rate: 30 fps Tiny YOLOv2 *Note: 540P means 960x540 ISP API Graphics API GRP Draws bounding box & label Get VGA resolution image Other reference designs available at: avnet.me/RzBoard-V2L © 2022 Avnet
- 9. Lightning-fast Stereo Vision on the Edge AI Kit
- 10. What is the i.MX 8M Plus Edge AI Kit? SMARC Carrier and SoM • NPU (2.3 TOPS) • Image Signal Processor 12 MP @ 30 fps • Quad Core Arm Cortex-A53 ARM Applications Processor • Dual image signal processors (ISP) and two camera inputs 10 Dual camera adapter • Two onsemi AR1335 IAS Camera Modules • onsemi AP1302 ISP Applications • Stereo Vision • Object detection • Facial recognition • Image segmentation © 2022 Avnet
- 11. • eIQ is NXP’s ML software stack for i.MX (included as Yocto layer) • Delegates choose which operations to accelerate i.MX 8M Plus Edge Kit — eIQ + NPU 11 ADD ADD MUL MUL ADD D2 D1 MUL MUL Support for: • ArmNN • TensorFlow Lite • ONNX Runtime • PyTorch • OpenCV © 2022 Avnet
- 12. i.MX 8M Plus Edge AI Kit — Stereo Vision Demo 12 © 2022 Avnet Other reference designs available at: avnet.me/imx8mplus-edgeai IAS Camera module Dual Camera Adapter ISP IAS Camera module I.MX8M+ SMARC EP5 Carrier I.MX8M+ SMARC SoM NPU Cortex-M7 10.1” Display LVDS Cortex-A53 GC7000UL GPU MIPI-CSI eIQ® Resolution: 1080P Frame Rate: 30 fps Display Picture Size：1280 x 800px Input Image at 30 fps
- 13. Low-cost, High-performance Machine Vision with the ZUBoard 1CG
- 14. • Versatile sub-$150 development board based on (AMD) Xilinx ZU1 Zynq UltraScale+ MPSoC • Dual arm A53 APU and arm R5 RPU • 256 KB on-chip RAM • 81.9 K FPGA Logic Cells • Similar to Avnet’s Zynq-7000 MiniZed • Dozens of ML models available via Vitis-AI • Sensors: Temperature & Pressure • Expansion Interfaces: 3 Syzygy, 1 Click What is ZUBoard 1CG? 14 Available 3Q22 Applications • Object detection • Facial recognition • Machine vision • Stereo vision © 2022 Avnet
- 15. Face Detection with the ZUBoard 1CG 15 © 2022 Avnet Other reference designs available at: avnet.me/ZUBoard_1CG ZUBoard 1CG GB Ethernet HD Webcam Resolution: 720P Processing system Cortex- A53 Cortex- R5F Web interface 640x480px @ 10 fps USB Programmable logic Output VGA resolution image
- 17. Benchmarking Results — Performance 17 0 50 100 150 200 250 300 350 400 ResNet50v1 TinyYolo v3 MobileNet v1 MobileNet v2 VGG16 Inference Time (milliseconds) RzBoard (DRP-AI) Edge AI Kit (NNAPI) ZUBoard 1CG © 2022 Avnet
- 18. Benchmarking Results — Power 18 0 0.5 1 1.5 2 2.5 3 ResNet50v1 TinyYolo v3 MobileNet v1 MobileNet v2 VGG16 Peak Power Consumption (Watts) RzBoard Edge AI Kit ZUBoard 1CG © 2022 Avnet
- 19. RzBoard V2L Conclusion 19 8M Plus Edge AI Kit ZUboard 1CG Applications • Object detection • Facial recognition • Machine vision • Stereo vision Applications • Stereo vision • Object detection • Facial recognition • Image segmentation Applications • Barcode scanning • Object detection • Pose detection • SLAM avnet.me/imx8mplus-edgeai avnet.me/RzBoard-V2L avnet.me/ZUBoard_1CG © 2022 Avnet
- 20. Resources Reference Designs: Edge AI Kit: http://avnet.me/imx8mplus-edgeai RzBoard: http://avnet.me/RzBoard-V2L ZUboard 1CG: http:// avnet.me/ZUBoard_1CG Board Info: Edge AI Kit: http://avnet.me/imx8mplus-edgeai RzBoard: http://avnet.me/RzBoard-V2L ZUboard 1CG: http://avnet.me/ZUBoard_1CG 20 © 2022 Avnet
- 21. Backup Slides
- 22. Embedded Vision Summit Avnet Booth Demos # Avnet SBC Board Featured Camera Solution Camera Spec Supplier Focus Demo Description 1 XBoard ZU1 SYZYGY dual IAS camera pod with AR0144 IAS camera modules (2) AP1302 ISP co-processor 1 MP (2x) Machine Vision / Stereo face-detection (Social-distancing health-safety application?) 2 8M Plus Edge AI Kit Dual IAS camera adapter with AR1335 IAS camera modules (2) AP1302 ISP co-processor 13 MP (2x) Facial recognition 3 RzBoard V2L Pi IAS camera adapter AS0260 IAS camera module 2 MP DRP-AI based object detection + comparison vs non DRP-AI (Smart Retail application?) 4 IAS Camera Modules AR0144 IAS camera module AR1335 IAS camera module AS0260 IAS camera module 1 MP 13 MP 2 MP Static demo with PowerPoint IAS overview © 2022 Avnet 22
- 23. Benchmarking Results — Latency (msec) 23 Model RzBoard Edge AI Kit ZUBoard 1CG Inception V4 N/A 101.41 194.64 VGG-16 N/A 82.50 206.83 ResNet-50 60.2 36.52 57.76 Mobilenet-V1 23.4 7.78 12.35 Mobilenet-V2 28.5 9.09 11.25 Tiny Yolo V3 41.2 21.07 44.97 SSD Mobilenet-V1 N/A 20.02 28.81 SSD Mobilenet-V2 N/A 36.32 49.50 © 2022 Avnet