Running AI on the edge
Nick Destrycker
Why Edge Computing?
Where is it used?
How AI works?
AI Hardware
Edgise
Table of contents
Edgise
Why edge computing ?
Edgise
Why edge computing ?
Edgise
H o w m u c h d a t a ?
1 2 8 0 x 7 2 0
= 9 2 1 6 0 0 P i xe l s
Why edge computing ?
Edgise
H o w m u c h d a t a ?
1 2 8 0 x 7 2 0
= 9 2 1 6 0 0 P i xe l s
1 P i xe l = 2 4 b i t s
Why edge computing ?
Edgise
H o w m u c h d a t a ?
1 2 8 0 x 7 2 0
= 9 2 1 6 0 0 P i xe l s
1 P i xe l = 2 4 b i t s
= 2 2 1 1 8 4 0 0 b i t s
Why edge computing ?
Edgise
H o w m u c h d a t a ?
1 2 8 0 x 7 2 0
= 9 2 1 6 0 0 P i xe l s
1 P i xe l = 2 4 b i t s
= 2 2 1 1 8 4 0 0 b i t s
Wh a t d o we l o o k f o r ?
2 s e a t s a r e t a ke n
- > r e p r e s e n t a b l e b y :
2 b i t s
Why edge computing ?
Edgise
H o w m u c h d a t a ?
1 2 8 0 x 7 2 0
= 9 2 1 6 0 0 P i xe l s
1 P i xe l = 2 4 b i t s
= 2 2 1 1 8 4 0 0 b i t s
Wh a t d o we l o o k f o r ?
2 s e a t s a r e t a ke n
- > r e p r e s e n t a b l e b y :
2 b i t s
2 2 1 1 8 3 9 8 b i t s o r
9 9 , 9 9 9 9 9 1 % o v e r h e a d
Reliability
Low latency
Privacy
Efficient use of bandwidth
Networking cost reduction
Edge approach
High bandwidth use
No connection, no decision
High and unpredictable latency
Classic approach
Why edge computing ?
Edgise
Why edge computing ?
Edgise
Smart City
MonitoringGai n val uabl e i nsi ghts usi ng edge computi ng camera anal ysi s
cl assi fyi ng several objects and scenes throughout the ci ty
S M A R T O B J E C T / S C E N E D E T E C T I O N
L O W P O W E R / N E T W O R K T R A F F I C
D ATA P R I VA C Y M A I N TA I N E D
Factory Automation
Recogni ze features of products on conveyer bel ts and actuate the
producti on l i ne accordi ngl y
A I V I S I O N S Y S T E M
D E T E C T / C L A S S I F Y P R O D U C T F E AT U R E S
H I G H S P E E D O F F L I N E P R O C E S S I N G
Human Presence
Detection
make sure autonomous vehi cl es stop when humans wal k i n front of
them
R E A C T I O N I N M I L L I S E C O N D S
L O W P O W E R / N E T W O R K T R A F F I C
N O H I G H S P E E D C O N N E C T I O N N E C E S S A R Y
Sound anomaly
detection
Detect sound anomal i es, not onl y rel ated to vol ume, but al so to the
nature of the sounds.
I N H E R E N T P R I VA C Y – S O U N D N O T R E C O R D E D
C O M P L E X A N A LY S I S
L O W P O W E R / N E T W O R K T R A F F I C
How does it work ?
How neural networks work
Edgise
How neural networks work
Edgise
Output = F(X0.W0 + X1.W1 + X2.W2 + X3.W3 … + Xn.Wn)
4 multiplications
3 summations
It’s all about MACs !
Detecting digits
Edgise
Detecting digits
Edgise
AI Hardware
Edgise
CPU
Edgise
GPU
Edgise
ASIC
Edgise
Performance
1 FPS
Power consumption
4,6mW
Edgise
Performance example
Low power
Performance
9000 FPS
Power consumption
11.7W
Hight performance
VGG-3, 32x32 images
2.15 x 2.55 mm
VGG-6, 32x32 images
17.0 x 17.0 mm
Bug FixingExpensive Time-To-Market
Hardware Development
Edgise
Idea to prototype
Full Package
Flexible development path
No vendor lock-in
Tailored Hardware
Adopt the software
way of working
Agile Development
The Edgise Way
Edgise
Optimal performance
Low unit cost
ASIC
Short Time-To-Market
Low development cost
Off-the-shelf
Increased reusability
Increased flexibility
Reconfigurable FPGA
Tailored Hardware
Edgise
WORK
FLOW
BACKLOG
REFINEMENT
SPRINT PLANNING
Using this approach we can address complex
adaptive problems, while productively and
creatively delivering products of the highest
possible value.
SHIPPABLE PRODUCT
DAILY
STANDUP
Edgise
DAILY STANDUP
• Review progress from last day
• Define stakeholder contact points
SPRINT REVIEW
• Discuss project progress
• Finetune scope
• Discuss issues
• still on budget ?
Every 2 weeks (or when needed)
Sprint Approach
Get Connected
https://www.edgise.com
Gaston Geenslaan 11B4, Leuven
hello@edgise.com
+32 472 43 95 42
+32 477 33 24 63
3000 Leuven
België
Gaston Geenslaan 11B4
Sam Sterckval
sam@edgise.com
+32 477 33 24 63
Nick Destrycker
nick@edgise.com
+32 472 43 95 42

Openbar Leuven // Edge-Computing: On-device AI // Nick Destrycker

  • 1.
    Running AI onthe edge Nick Destrycker
  • 2.
    Why Edge Computing? Whereis it used? How AI works? AI Hardware Edgise Table of contents Edgise
  • 3.
  • 4.
    Why edge computing? Edgise H o w m u c h d a t a ? 1 2 8 0 x 7 2 0 = 9 2 1 6 0 0 P i xe l s
  • 5.
    Why edge computing? Edgise H o w m u c h d a t a ? 1 2 8 0 x 7 2 0 = 9 2 1 6 0 0 P i xe l s 1 P i xe l = 2 4 b i t s
  • 6.
    Why edge computing? Edgise H o w m u c h d a t a ? 1 2 8 0 x 7 2 0 = 9 2 1 6 0 0 P i xe l s 1 P i xe l = 2 4 b i t s = 2 2 1 1 8 4 0 0 b i t s
  • 7.
    Why edge computing? Edgise H o w m u c h d a t a ? 1 2 8 0 x 7 2 0 = 9 2 1 6 0 0 P i xe l s 1 P i xe l = 2 4 b i t s = 2 2 1 1 8 4 0 0 b i t s Wh a t d o we l o o k f o r ? 2 s e a t s a r e t a ke n - > r e p r e s e n t a b l e b y : 2 b i t s
  • 8.
    Why edge computing? Edgise H o w m u c h d a t a ? 1 2 8 0 x 7 2 0 = 9 2 1 6 0 0 P i xe l s 1 P i xe l = 2 4 b i t s = 2 2 1 1 8 4 0 0 b i t s Wh a t d o we l o o k f o r ? 2 s e a t s a r e t a ke n - > r e p r e s e n t a b l e b y : 2 b i t s 2 2 1 1 8 3 9 8 b i t s o r 9 9 , 9 9 9 9 9 1 % o v e r h e a d
  • 9.
    Reliability Low latency Privacy Efficient useof bandwidth Networking cost reduction Edge approach High bandwidth use No connection, no decision High and unpredictable latency Classic approach Why edge computing ? Edgise
  • 10.
  • 11.
    Smart City MonitoringGai nval uabl e i nsi ghts usi ng edge computi ng camera anal ysi s cl assi fyi ng several objects and scenes throughout the ci ty S M A R T O B J E C T / S C E N E D E T E C T I O N L O W P O W E R / N E T W O R K T R A F F I C D ATA P R I VA C Y M A I N TA I N E D
  • 12.
    Factory Automation Recogni zefeatures of products on conveyer bel ts and actuate the producti on l i ne accordi ngl y A I V I S I O N S Y S T E M D E T E C T / C L A S S I F Y P R O D U C T F E AT U R E S H I G H S P E E D O F F L I N E P R O C E S S I N G
  • 13.
    Human Presence Detection make sureautonomous vehi cl es stop when humans wal k i n front of them R E A C T I O N I N M I L L I S E C O N D S L O W P O W E R / N E T W O R K T R A F F I C N O H I G H S P E E D C O N N E C T I O N N E C E S S A R Y
  • 14.
    Sound anomaly detection Detect soundanomal i es, not onl y rel ated to vol ume, but al so to the nature of the sounds. I N H E R E N T P R I VA C Y – S O U N D N O T R E C O R D E D C O M P L E X A N A LY S I S L O W P O W E R / N E T W O R K T R A F F I C
  • 15.
  • 16.
  • 17.
    How neural networkswork Edgise Output = F(X0.W0 + X1.W1 + X2.W2 + X3.W3 … + Xn.Wn) 4 multiplications 3 summations It’s all about MACs !
  • 18.
  • 19.
  • 20.
  • 21.
  • 22.
  • 23.
  • 24.
    Performance 1 FPS Power consumption 4,6mW Edgise Performanceexample Low power Performance 9000 FPS Power consumption 11.7W Hight performance VGG-3, 32x32 images 2.15 x 2.55 mm VGG-6, 32x32 images 17.0 x 17.0 mm
  • 25.
  • 26.
    Idea to prototype FullPackage Flexible development path No vendor lock-in Tailored Hardware Adopt the software way of working Agile Development The Edgise Way Edgise
  • 27.
    Optimal performance Low unitcost ASIC Short Time-To-Market Low development cost Off-the-shelf Increased reusability Increased flexibility Reconfigurable FPGA Tailored Hardware Edgise
  • 28.
    WORK FLOW BACKLOG REFINEMENT SPRINT PLANNING Using thisapproach we can address complex adaptive problems, while productively and creatively delivering products of the highest possible value. SHIPPABLE PRODUCT DAILY STANDUP Edgise DAILY STANDUP • Review progress from last day • Define stakeholder contact points SPRINT REVIEW • Discuss project progress • Finetune scope • Discuss issues • still on budget ? Every 2 weeks (or when needed) Sprint Approach
  • 29.
    Get Connected https://www.edgise.com Gaston Geenslaan11B4, Leuven hello@edgise.com +32 472 43 95 42 +32 477 33 24 63 3000 Leuven België Gaston Geenslaan 11B4 Sam Sterckval sam@edgise.com +32 477 33 24 63 Nick Destrycker nick@edgise.com +32 472 43 95 42