AI applications at the very edge loic.lietar@greenwaves-technologies.com www.greenwaves-technologies.com 1
About GreenWaves Technologies • Founded by industry veterans in November 2014 • Based in Grenoble area, France • 16 people, going to 30 • Launched first IoT application Processor, GAP8, in February 2018
Could that scale at IoT levels? Cloud computing + • Installation costs $$$ • Privacy Proprietary Information 3
Could that scale at IoT levels? Cloud computing Edge computing + + • Installation costs $$ • Bandwidth • Privacy Proprietary Information 4
Could that scale at IoT levels? Cloud computing Edge computing Very edge computing + + • Installation costs $$ • Bandwidth • Privacy Proprietary Information 5
This can scale Cloud computing Edge computing Very edge computing + + • Installation costs $ • Operation cost $ • Bandwidth • Privacy Proprietary Information 6
Is the HW ecosystem ready? • Power supply • Battery: 3.6V 3.6Ah A size battery, 2% loss per year i.e. 250uW avg for 5 years • PV: 10uW per cm 2 for 50 lux indoor (ceiling) (direct sun is 100k lux) i.e. 250uW avg for 50 cm 2 12 hours a day • Sensors • Wireless communication • Processing Proprietary Information 7
Sensors • QVGA, 30fps @ 2mW Se # frames • VGA global shutter with logarithmic uW mW image fps to snap an (1 image/min) sensitivity, 200 fps @ 230mW image QVGA 2 30 6 6,7 • IR 80x80, 10 fps @ 15mW VGA 230 200 1 19,2 • 10Hz 2m radar @ 1mW The numbers illustrate the potential of the technology power wise • System wise, the sensors architecture and features only start to be • • microphone @ 300uW designed for IoT use … e.g. some microphones now support wake-up-on-noise and low • resolution/lower power mode The devil is in the detail some image sensors offer multiple resolutions, but no image • sensor is meant for snapping one image at the time. Their use remains a bit of DIY Plenty of room for improvement • Proprietary Information 8
Wireless communication, not a simple story Proprietary Information 9
Processing • As for sensors, raw computing power energy efficiency is not enough for IoT • IoT devices have multiple states, each of which the processor shall match to consume just the minimum necessary energy GAP8 (up to 11GOPS) • Deep sleep 1uA • Data acquisition 50uA • System control/light DSP 3 to 10 mW • DSP like 20 to 80 mW • Specialized computing (CNN) + 4 to 14 mW uW (1 image/min) 10 objects recognition CNN 250 QVGA Face Detection Viola-Jones 8 GAP8 Pedestrian Detection Weak predictors 47 10 objects recognition CNN 1250 VGA GAP9 10 objects recognition CNN 313 Proprietary Information 10
In conclusion we can support an event from once every few minutes to few times per minute, on average, for 5 years (battery) or “for ever” (PV cell) uW Power 250 once a minute image sensor 7 10 objects recognition 250 QVGA Face Detection 8 Pedestrian Detection 47 image sensor 19 VGA 10 objects recognition 1250 10 kbit/s 24 LoRa 1 kbit/s 240 Proprietary Information 11
In real life • To minimize energy consumption, we also go for hierarchical device architecture, from always-on inaccurate low power sensor to higher power higher resolution sensor • e.g. • a PIR sensor detects life heat • the camera is turned on during the day (an IR camera during the night) and captures an image • the processor looks for a person in the image • a basic microphone detects a sound • the microphone array is turned on and captures an audio sequence • the processor calculates the direction of the sound and defines the window of interest • the camera is turned and captures an image • the processor looks into the window of interest for what the algorithm as been trained for Proprietary Information 12
But are there uses for very edge computing? ≠ HD video stream QVGA to 1Mpixel images (and others) Orientable Fixed Unlimited computing power Limited computing power Proprietary Information 13
Applications Counting cars, bicycles Preventive maintenance, Spotting stopped cars on the road side monitoring sound and vibrations License plate reading Detecting face Traffic classification Face recognition (few, many) Abnormal noises identification Counting people in meeting room Retail, detecting pedestrians Estimating occupancy in cafeteria Hotel room entrance monitoring Proprietary Information 14
Applications Vital signs monitoring Intrusion classification Window shock classification Gesture recognition Key words spotting Inattention detection Proprietary Information@ 15
Thank you www.greenwaves-technologies.com 16
Recommend
More recommend
