review of factors affecting IoT system selection. for MVP phase and later phases. Computation, price, connectivity, open source support, development SDKs
2. Outline
• IoT Modern needs
• Deep Learning networks and operations
• CPU GPU TPU and NN
• Considerations
• Dev-boards and NPU accelerators
– RPI
– NVIDIA
– Rockchip
– Intel
• Cameras 2D / 3D
• Summary
4. Basic Edge Processor operation
• Worked as controller
• Control motors / open gates
• Take pictures / basic image operations
• Send images to the cloud for further
processing/storage/decision
• Operate based on cloud decision
6. Neural Networks
• Basic idea: Multiply each input by weights
and sum
• Exists for 60+ years !!
– Implemented with resistors as weights
• Deep Neural networks 100+ layers
– Hundreds of millions or more weights
– Heavy matrix operations
7. LSTM/RNN
• Used for
– sound detectors (gun shot detector)
– data series - number of clients in restaurant
– Activity detector – based on accelerometers
– Usually requires low processing compared to
CNNs
8. CNN - Classifiers
• Used for
– Is object X in the image
– Is there a cat in the Image ?
– Common implementation LaNet
9. BOX Detectors
• Box detectors detects where is the object
located
• Used to detect:
– People, face, corona mask
– car, road signs
– cats
• Known box detectors:
– YOLO3 YOLO5
– SSD
10. Segmenter and Multi-Instance Segmenter
• Segmentor
– defines class for each pixel
– Pixel(X,Y) belongs to cat
– FCN-8 basic implementation
– Does not separate cats
• Multi-instance segmentor
– Separate Cats
– Pixel(X,Y) belongs to cat42
– MRCNN – common implementation
11. CPU and CNNs
• X86, ARM – sequential instruction
processing
• Not suitable for matrix (DL) operations !
• Remedies
• +SIMD – MMX SSE
• +Multicore 2,4,8….
• Yet….
13. GPU – more than a graphic processor
• GPU are using 100s-1000s of (simple)
cores
• Better in parallel processing than CPU
• Much better for large matrix operations
13
14. GPU Examples
• GPU Cards
– NVIDIA 20XX
– NVIDIA T4
• Can we run a GPU only IoT solution?
15. TPUs
• TPUs are built ONLT for deep learning
• Designed for single operation layer
processing
• First GPUs started 4 years ago
• Google Coral, Hauwei, NVIDIA, Intel
Movidius, Narvana,
• In Israel: Habana, Heilo
• In most modern phones: Samsung apple,
QCOM
19. Edge TPU accelerators
• Intel Movidius Processor
– Edge AI processor / accelerator
– Either chip or USB
– Connects to RPI
• Google TPU (Coral)
– Chip or USB
• Rockchip RK 1808
20. Processors today are SoC
• SoC which usually include:
– 4-8 ARM cores
– GPU (ARM Mali, NVIDIA XX, Intel XX)
– Sometimes a TPU
• The selection is CPU/GPU/TPU
• Selection of solution is between SoC
• For MVP or development phases selection is
between SoC development boards !!
21. Considerations
• Price - $$
• Performance TFlops/Sec
• Power - Watts
• Longevity
• Existing code base
• Development boards
• Connectivity
• SDKs
29. Intel Atom
• UP – boards
• 200-700$ depending on kit including
RealSense for some options
• Use OpenVino SDK for development and
optimization
30. Jetson NX
• 399$ dev board
• X4-5 stronger than the nano
• Same RPI-like form factor
and connectivity
• emmc + PCIe
• Can do:
31. Software / open source
• RPI has many open-source solutions
• NVIDIA uses DeepStream and ISAAC SDKs
to accelerate development
• Intel uses OpenVino to accelerate
development on its boards
32. 2D Camera
• MIPI / USB connectivity / HDMI
• Resolution / Formats YUV, JPEG, Bayer
• HDR / Pixel size
• Filters, IR Cut
• Lens quality, FOV, distortion
• ISP quality – very important for DL
• Not all cameras supported by all boards
• USB camera – UVC no need for driver on
Linux windows
33. Placing AI in the camera
• Many IoT solutions require depth camera
/ Lidar - Lots of processing
• To reduce processor cost done inside the
camera:
– Intel Realsense family
– Rockchip RMSL 2031000
– Veronica - Inutive
34. RealSense
• D455/D435/D415 cameras
– IR + RGB + Projector + dual cam
• Option for IMU for robots/drones
• On board processor
• Not fit for RPI in high-res(high data rate)
35. Inutive - Veronica
• Designed for robots drones AR
• Low latency to avoid AR motion sickness
• On chip SLAM
• Built in VPU
• Output depth map
36. ROCKchip RMSL203
• Rockchip depth camera solution
• Built in VPU/DSP
• Output depthmap
• Projector + IR + RGB
37. Placing AI in the camera
• Many IoT solutions require depth camera
/ Lidar - Lots of processing
• To reduce processor cost done inside the
camera:
– Intel Realsense family
– Rockchip RMSL 2031000
– Veronica - Inutive
38. RealSense
• D455/D435/D415 cameras
– IR + RGB + Projector + dual cam
• Option for IMU for robots/drones
• On board processor
• Not fit for RPI in high-res(high data rate)
39. Inutive - Veronica
• Designed for robots drones AR
• Low latency to avoid AR motion sickness
• On chip SLAM
• Built in VPU
• Output depth map
40. ROCKchip RMSL203
• Rockchip depth camera solution
• Built in VPU/DSP
• Output depthmap
• Projector + IR + RGB
41. Summary
• There isnt a wining solution
• Robotics / drones solutions differ from
video analytics
• First product MVP should prefer
development time & costs over
price/performance
• Select a product with good, well tried
SDKs and community
42. Cloud connectivity
• Why we need it?
– Heavy data/image/deep learning processing
– Data aggregation from many sensing points
– Logging
– Dashboard
– Health test
– Secure access
– OTA updates
• Both AWS and Azure has Cloud IoT solution which
seamless integrate with RPI3 and Jetson Nano
43. NVIDIA and AWS IoT (Core & Greengrass)
• Connecting to AWS IoT is easy
• Deepstream has already built in MQTT message
formatters and brokers designed for AWS
connectivity
Source: https://aws.amazon.com/blogs/iot/how-to-integrate-nvidia-deepstream-on-jetson-devices-with-aws-iot-core-and-aws-iot-
greengrass/
44. Cloud connectivity
• Why we need it?
– Heavy data/image/deep learning processing
– Data aggregation from many sensing points
– Logging
– Dashboard
– Health test
– Secure access
– OTA updates
• Both AWS and Azure has Cloud IoT solution which
seamless integrate with RPI3 Jetson products and
Intel products
45. Summary
• There isn't a wining solution
• Robotics / drones solutions differ from
video analytics
• First product MVP should prefer
development time & costs over
price/performance
• Select a product with good, well tried
SDKs and community
• Invest in a good camera / ISP
46. Contact
• Yossi Cohen CTO@DSP-IP
• +972-54-5313092 | yossi@dsp-ip.com
• What we do:
– Video Analytics
– Robotics
– Drones
– Autonomous cars
– Agritec
Editor's Notes
We will see how the migration affects each of those steps
This requires a large investment in server side software, However cloud vendors mapped all those requirements and create a generic platform which supports them
This requires a large investment in server side software, However cloud vendors mapped all those requirements and create a generic platform which supports them