Edge computing is set to be the next big thing in the Internet of Things. There is a rush to bring new low power AI chips to market which will allow devices to have greater autonomy, particularly in areas like video processing and self-driving vehicles. This presentation looks at the current state of development and the emergence of a new generation of silicon startups.
the rest of
Amount of time spent designing products
Amount of time spent discussing AI
The Comms part is largely done
Low cost data exists
Just because it will get better is not a
reason for prevarication.
Data Capture Data Insight
There’s a lot of detail in between…
The IoT value stack
Deployment & Physical installation
Additional Data Sourcing
Business Applications (vertical)
Business Applications (packaged)
Data Cleansing & Verification
Security & Updates
Sensor & Physical
M2M / IoT
(DLC – Device
The world is producing excessive amounts of
“unstructured data” that need to be reconstructed.
Rob High – CTO, IBM
Big Data doesn’t need to reside in one place.
Lots of Little Data is also Big Data.
Learning can be distributed.
Because Intel wants to sell more server chips.
Because CISCO wants to sell more infrastructure.
Because the network operators need a story to support 5G.
Why is edge computing such a well kept secret?
And also because it’s difficult.
The balance of power
• Limited processing power
• Limited resources
• Limited battery life
• Intermittent connectivity
• Lots of processing power
• Lots of resources
• Mains powered
• Aggregated Data
• Additional Data Sources
The balance of power
• May need to make real time
• Can’t guarantee a connection
• May have limited data throughput
• Intermittent uploads
• Very limited downloads
• Little access to additional data
• Difficult to make real-time control
decisions for millions of devices
The processing hierarchy
• Heavy Lifting
• “Unlimited” resources
• Pre-programmed and
• Video processing, etc.
• Real-time learning
• Autonomous operation
Giga (Billion) Operations per second and Trillion Operations per second
TOPS and GOPS
Intel Xeon 8180M 0.3 TOPS / W
NVIDIA 0.4 TOPS / W
< 0.05 TOPS 2 - 3 TOPS 25 - 50 TOPS
GreenWaves 0.6 TOPS / W
Kneron offers 1.5 TOPS / W
ARM ML 3 TOPS / W
Novumind 3 TOPS / W
Cambricon 3 TOPS / W
Mythic 4 TOPS / W
Groq 8 TOPS / W
Syntiant 20 TOPS / W
Is it training or is it inference?
MLP - Multi-layer Perceptron
CNN - Convolutional Neural Networks
RNN - Recurrent Neural Networks
DNN - Deep Neural Networks – image recognition & voice
The AI Landscape
Video Neural Network Engines and AI accelerators
Sunrise AI chip for Facial Recognition
Supports 4 x 1920 x 1080 30fps video inputs at under 1.5W
Google’s Edge TPU
“Edge-based ML inference is vital to delivering reliable,
live, low-latency, and cost-effective smart city IoT. Cloud
IoT Edge and Edge TPU unlock these capabilities in new
ways for the next generation of Smart Parking systems.”
John Heard, Chief Technology Officer, Smart Parking Limited
Edge TPU Features
“The first step in a roadmap that will leverage Google's AI expertise
to follow and reflect in hardware the rapid evolution of AI.”
• Inference Accelerator
• Dev boards coming soon