SlideShare ist ein Scribd-Unternehmen logo
1 von 22
Downloaden Sie, um offline zu lesen
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
An Experimental Implementation of
an Edge-based AI Engine
with Edge-Cloud Coordination
ISCIT-2018, Bangkok, Thailand
Toshihiko Yamakami
CTO Team, ACCESS
Toshihiko.Yamakami@access-company.com
2018/09
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 1 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Outline
Background
Related Work
Research Method
Design
Edge and Cloud Coordination Requirements
Architecture
Implementation
Prototype Implementation
Model Switch Performance
Discussion
Advantages of the Proposed Model
Possible Mechanism for Stability
Limitations
Conclusion
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 2 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Background
Research Purpose
The aim of this research is to develop device-edge-cloud coordination in
edge AI chip-based intelligent IoT systems.
Background
Cloud-based AI has advantages of flexibility, capacity and throughput,
but it has disadvantages of high communication costs and long delays.
Edge represents a node which is connected to devices and a cloud. It
has low latency direct connection to devices. It also has a connection
to a cloud, but does not have the full capabilities of the cloud,
It is promising to utilize an edge AI development environment where a
framework of edge-cloud AI coordination is deployed.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 3 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Related Studies
Research on edge intelligent computing consists of three areas:
a) feasibility and framework of edge computing: fine-grained edge offloading
architecture [Cozzolino17], auction scheme in edge resource allocation
[Prasad17], conditions for edge computing feasibility [Kalim17]
b) AI in edge computing: latency-aware edge computing in video
analytics[Yi17], distributed processing of a Deep CNN in smart camera
[Castillo17], feasibility of AI and ML at edge computing [Sharma18]
c) intelligent systems in edge-cloud coordination: healthcare optimization in
wearable-edge-cloud coordination[Strässle17] docker overhead in Mobile
Edge Computing [Avino17]
The past research did not cover the special purpose AI-engine-chip-based
device-edge-cloud coordination from the software perspective.
The originality of this paper lies in its identification of software-based
approach of AI-engine-chip-based device-edge-cloud coordination.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 4 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Research Method
Identifying Edge AI requirements,
Designing a Prototype Implementation of Edge-Cloud Coordinated AI,
Implementing a Prototype using FPGA (Field-Programmable Gate Array)
Discussing Lessons Learned in Prototype Implementation
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 5 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Challenges of AI-engine-based AIoT systems
Aspect Issues
AI-Engine Induction performance, Power-consumption-efficiency,
Low-latency, Privacy of data, Cost
Cloud Learning efficiency, Performance of learning models
Edge-Cloud coordi-
nation
Alignment of edge processing and cloud processing,
Controllability of device execution using knowledge in
the cloud.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 6 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Requirements of Edge AI Computing
Item Description
Low latency Intelligent device control requires low latency. It is dif-
ficult at cloud-based AI systems.
Low communication
costs
Massive IoT systems need low communication costs in
order to prevent cost explosion. The communication
costs to the cloud can be a prohibitive factor.
Privacy and security In-edge AI computing does not expose data to be pro-
cessed. It is difficult at cloud-based AI systems.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 7 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Advantages of edge-cloud coordination
Item Description
Utilization of Data
Analytics
The edge-cloud coordination enables utilization of
stored learned models at the cloud.
Coordination of dis-
tributed edges
The edge-cloud coordination enables real-time context
utilization with aggregated data from distributed edges
in a real-time manner.
Task migration The edge-cloud coordination enables task migration
and resource optimization among multiple edges.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 8 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Use cases of Edge-Cloud Coordinated AI
Item Description
Data sharing When necessary, data at the edge need sharing with the
cloud.
Cloud-based control When a context is recognized by a cloud, a cloud-based
control needs coordination at the edge side so that an
intelligent cloud-based control is facilitated. For exam-
ple, learned models can be switched according to the
request from the cloud.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 9 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
A prototype environment
Item Description
Simulating environ-
ment
Xilinks ep706-based DMP evaluation SDK kit.
CPU FPGA embedded
OS Xilinks Linux OS
Simulated core DMP IP Core
Languages C++ to control device logics, Python to interact with
a cloud.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 10 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Architecture Overview
Device
(Camera,
sensors, …)
Edge
(AI engine)
Cloud
(storage,
management)
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 11 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Edge Architecture
High-level Modules
AI Module
Induction
Module
Learned
Model
Manager
Cloud
Coordination
Module
Low-level Modules
Communication
Module
Local Storage
Module
Camera
Module
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 12 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Implementation
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 13 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Demo features implemented on a prototype environment
Item Description
Video recording Video recording is performed with a camera connected
to the device.
Object detection Objects are detected by the pre-set learned models. The
number of detected objects are counted and compared
with the pre-set threshold.
Image sharing Images with rectangle-framed detected objects are up-
loaded.
Remote setup of
learned models
The learned models can be switched by a command
from a cloud using a switch-and-reset command.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 14 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Model Switch Performance
This paper does not cover detailed performance analysis of the current
implementation.
For model switch time, It takes approximately 15 second to switch learned
models In the current implementation. The major time consumption lies in
a) model load and execution and b) waiting time for graceful exit for model
loading. Without the latter, the execution of new learned model is unstable.
Also, the acknowledgement of model switch command takes 3 seconds at
maximum.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 15 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Advantages of the Proposed Approach
The implementation of a prototype in FPGA-simulated IP core verifies
points (see Result Table)
The demo demonstrates the feasibility of edge-cloud coordination
using ASIC-based AI engines as edge computing.
There are lessons learned in the implementation on needs of
software-based support for hardware-involved integration (see Lessons
Table).
Possible mechanisms to enhance stability of edge AI computing are
needed (see Possible mechanisms table).
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 16 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Results of the implementation
Item Description
Edge-closed process-
ing
Low latency, privacy, and communication-less process-
ing are implemented.
High-level learned
model management
It is feasible to switch multiple learned models depend-
ing contexts.
On-demand infor-
mation sharing
When a context requires full image transfer, needed
images can be uploaded to the cloud.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 17 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Lessons learned in the implementation
Item Description
Timing to ASIC It is critical to align some hardware timing. When the
timing is not aligned, a switch of learned models does
not work and the system is frozen without any outputs.
Timing coordination
between an edge and
a cloud
In an object detection use case, time alignment between
an edge and a cloud is subtle. The slight time differ-
ence gives a small discrepancy between the detected
rectangles and the transferred image.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 18 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Possible mechanisms to enhance stability of edge AI computing
Item Description
Debug Logging, tracing, and probing functions are provided.
Self-diagnosis Self diagnosis of monitoring process is performed in or-
der to identify induction process faults without human
intervention..
Adaptive computing Fault-tolerant processing when some parts of camera-
AI process chain do not work properly.
Remote setup of
learned models
The learned models can be switched by a command
from a cloud using a switch-and-reset command.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 19 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Limitations
This implementation is a proof of concept implementation. Camera works
and detected data can be shared, but performance, scalability, and reliability
are out of the scope of this paper.
Function is based on conceptual verification. Command protocol is not
implemented and interaction with the cloud is performed by the state
variable control at the cloud. Images are not encoded at the device and
shared by a binary array format. In-depth synchronization of detected
objects and shared images remains for future studies.
This research is a prototype implementation based on a simulated
environment constructed on ASIC. The real world deployment will be based
on specialized battery-efficient Edge AI processor core. The system
performance for such specialized AI hardware remains future studies.
Theoretically, model conversions from Caffe should work, but it does not
perform reliably on the implementation in this paper.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 20 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Conclusion
Requirements of low latency, privacy, and limited bandwidth drive the shift
from cloud computing and edge computing.
In this paper, the design of an experimental prototype for device-edge-cloud
coordination for object detection use cases is presented. The feasibility of
device-edge-cloud coordination is verified by an implementation using FPGA.
It shows the effectiveness of edge-based real-time object detection with
high-level cloud-base manipulation of multiple learned models stored in an
edge node. This experimental implementation can lead to the full
ASIC-based implementation when the IP core is ASIC-fabricated.
At the same time, the experiment exposes challenges in hardware-software
coordination. The author discusses design of software-based compensation
towards these challenges.
The design derived from this implementation is a stepping stone for full
ASIC-based AI-engine solutions in the device-edge-cloud coordination.
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 21 / 22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Thank you for your attention!
Questions?
Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 22 / 22

Weitere ähnliche Inhalte

Ähnlich wie An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordination

Indoor Point Cloud Processing - Deep learning for semantic segmentation of in...
Indoor Point Cloud Processing - Deep learning for semantic segmentation of in...Indoor Point Cloud Processing - Deep learning for semantic segmentation of in...
Indoor Point Cloud Processing - Deep learning for semantic segmentation of in...
CubiCasa
 
Mohan_Dissertation (1)
Mohan_Dissertation (1)Mohan_Dissertation (1)
Mohan_Dissertation (1)
Mohan Bhargav
 
BP_Kucera_Adam_2016
BP_Kucera_Adam_2016BP_Kucera_Adam_2016
BP_Kucera_Adam_2016
Adam Ku?era
 

Ähnlich wie An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordination (20)

Indoor Point Cloud Processing - Deep learning for semantic segmentation of in...
Indoor Point Cloud Processing - Deep learning for semantic segmentation of in...Indoor Point Cloud Processing - Deep learning for semantic segmentation of in...
Indoor Point Cloud Processing - Deep learning for semantic segmentation of in...
 
Indoor Point Cloud Processing
Indoor Point Cloud ProcessingIndoor Point Cloud Processing
Indoor Point Cloud Processing
 
S3 Server Hackathon Presented by S3 Server, a Scality Product, Seagate and Ho...
S3 Server Hackathon Presented by S3 Server, a Scality Product, Seagate and Ho...S3 Server Hackathon Presented by S3 Server, a Scality Product, Seagate and Ho...
S3 Server Hackathon Presented by S3 Server, a Scality Product, Seagate and Ho...
 
Hackathon scality holberton seagate 2016 v5
Hackathon scality holberton seagate 2016 v5Hackathon scality holberton seagate 2016 v5
Hackathon scality holberton seagate 2016 v5
 
SparkEdgeEmu: An Emulation Framework for Edge-enabled Apache Spark Deployments
SparkEdgeEmu: An Emulation Framework for Edge-enabled Apache Spark DeploymentsSparkEdgeEmu: An Emulation Framework for Edge-enabled Apache Spark Deployments
SparkEdgeEmu: An Emulation Framework for Edge-enabled Apache Spark Deployments
 
Mohan_Dissertation (1)
Mohan_Dissertation (1)Mohan_Dissertation (1)
Mohan_Dissertation (1)
 
Project final report
Project final reportProject final report
Project final report
 
Brandtzaeg master
Brandtzaeg masterBrandtzaeg master
Brandtzaeg master
 
BP_Kucera_Adam_2016
BP_Kucera_Adam_2016BP_Kucera_Adam_2016
BP_Kucera_Adam_2016
 
Deploying Deep Learning Algorithm On AWS Cloud Platform.pdf
Deploying Deep Learning Algorithm On AWS Cloud Platform.pdfDeploying Deep Learning Algorithm On AWS Cloud Platform.pdf
Deploying Deep Learning Algorithm On AWS Cloud Platform.pdf
 
Modeling and Simulation of Parallel and Distributed Computing Systems with Si...
Modeling and Simulation of Parallel and Distributed Computing Systems with Si...Modeling and Simulation of Parallel and Distributed Computing Systems with Si...
Modeling and Simulation of Parallel and Distributed Computing Systems with Si...
 
Fowa Miami 09 Cloud Computing Workshop
Fowa Miami 09 Cloud Computing WorkshopFowa Miami 09 Cloud Computing Workshop
Fowa Miami 09 Cloud Computing Workshop
 
Portfolio_new
Portfolio_newPortfolio_new
Portfolio_new
 
Tigerstripe @ Eclipse Summit 08
Tigerstripe @ Eclipse Summit 08Tigerstripe @ Eclipse Summit 08
Tigerstripe @ Eclipse Summit 08
 
A Social Dimension View Model of Divergence of IoT Standardization
A Social Dimension View Model of Divergence of IoT StandardizationA Social Dimension View Model of Divergence of IoT Standardization
A Social Dimension View Model of Divergence of IoT Standardization
 
Madge Perspective
Madge PerspectiveMadge Perspective
Madge Perspective
 
IRJET- CFD-A Trend in Automobile Aerodynamics Technology
IRJET- 	  CFD-A Trend in Automobile Aerodynamics TechnologyIRJET- 	  CFD-A Trend in Automobile Aerodynamics Technology
IRJET- CFD-A Trend in Automobile Aerodynamics Technology
 
Object Detection for Autonomous Cars using AI/ML
Object Detection for Autonomous Cars using AI/MLObject Detection for Autonomous Cars using AI/ML
Object Detection for Autonomous Cars using AI/ML
 
Introduction to mago3D, an Open Source Based Digital Twin Platform
Introduction to mago3D, an Open Source Based Digital Twin PlatformIntroduction to mago3D, an Open Source Based Digital Twin Platform
Introduction to mago3D, an Open Source Based Digital Twin Platform
 
Sky X Tech Report
Sky X Tech ReportSky X Tech Report
Sky X Tech Report
 

Mehr von Toshihiko Yamakami

Mehr von Toshihiko Yamakami (20)

Toward Interactive People Analytics: A New Approach to Leverage Organizationa...
Toward Interactive People Analytics: A New Approach to Leverage Organizationa...Toward Interactive People Analytics: A New Approach to Leverage Organizationa...
Toward Interactive People Analytics: A New Approach to Leverage Organizationa...
 
Amazon SageMaker: 機械学習の民主化から工業化へ(in Japanese)
Amazon SageMaker: 機械学習の民主化から工業化へ(in Japanese)Amazon SageMaker: 機械学習の民主化から工業化へ(in Japanese)
Amazon SageMaker: 機械学習の民主化から工業化へ(in Japanese)
 
ホワイトペーパーの書き方 (in Japanese)
ホワイトペーパーの書き方 (in Japanese)ホワイトペーパーの書き方 (in Japanese)
ホワイトペーパーの書き方 (in Japanese)
 
認知科学に基づく「効果的なデモ」のヒント (in Japanese)
認知科学に基づく「効果的なデモ」のヒント (in Japanese)認知科学に基づく「効果的なデモ」のヒント (in Japanese)
認知科学に基づく「効果的なデモ」のヒント (in Japanese)
 
Singularity Literacy: Civic Skill Shift from Workplace to Studyplace
 Singularity Literacy: Civic Skill Shift from Workplace to Studyplace Singularity Literacy: Civic Skill Shift from Workplace to Studyplace
Singularity Literacy: Civic Skill Shift from Workplace to Studyplace
 
A Visualization Framework to Empower Small and Medium-Sized Enterprises in Op...
A Visualization Framework to Empower Small and Medium-Sized Enterprises in Op...A Visualization Framework to Empower Small and Medium-Sized Enterprises in Op...
A Visualization Framework to Empower Small and Medium-Sized Enterprises in Op...
 
行動誘導技術で読み解くアイディアマラソン (in Japanese)
行動誘導技術で読み解くアイディアマラソン (in Japanese)行動誘導技術で読み解くアイディアマラソン (in Japanese)
行動誘導技術で読み解くアイディアマラソン (in Japanese)
 
論文SEOとは何か?(in Japanese)
論文SEOとは何か?(in Japanese)論文SEOとは何か?(in Japanese)
論文SEOとは何か?(in Japanese)
 
A Dimensional Model of Service Design Toward Utilizing Public Transportation ...
A Dimensional Model of Service Design Toward Utilizing Public Transportation ...A Dimensional Model of Service Design Toward Utilizing Public Transportation ...
A Dimensional Model of Service Design Toward Utilizing Public Transportation ...
 
シニア・イノベーションの視点
シニア・イノベーションの視点シニア・イノベーションの視点
シニア・イノベーションの視点
 
An Organizational Coordination Model for IoT: A Case Study of Requirement Eng...
An Organizational Coordination Model for IoT: A Case Study of Requirement Eng...An Organizational Coordination Model for IoT: A Case Study of Requirement Eng...
An Organizational Coordination Model for IoT: A Case Study of Requirement Eng...
 
An Evolution-by-design Approach: Toward Multi-disciplinary Life-cycle Manage...
 An Evolution-by-design Approach: Toward Multi-disciplinary Life-cycle Manage... An Evolution-by-design Approach: Toward Multi-disciplinary Life-cycle Manage...
An Evolution-by-design Approach: Toward Multi-disciplinary Life-cycle Manage...
 
A Migration-oriented Partial Adaptation Architecture for IoT-empowered City P...
A Migration-oriented Partial Adaptation Architecture for IoT-empowered City P...A Migration-oriented Partial Adaptation Architecture for IoT-empowered City P...
A Migration-oriented Partial Adaptation Architecture for IoT-empowered City P...
 
A Dimensional Framework to Evaluate Coverage of IoT Services in City Platform...
A Dimensional Framework to Evaluate Coverage of IoT Services in City Platform...A Dimensional Framework to Evaluate Coverage of IoT Services in City Platform...
A Dimensional Framework to Evaluate Coverage of IoT Services in City Platform...
 
A Gap Analysis Framework of IoT-empowered City Platform as a Service
A Gap Analysis Framework of IoT-empowered City Platform as a ServiceA Gap Analysis Framework of IoT-empowered City Platform as a Service
A Gap Analysis Framework of IoT-empowered City Platform as a Service
 
デゲーミフィケーションとは何か ゲームの快感原理を学んで自己成長に活かす(in Japanese)
デゲーミフィケーションとは何か ゲームの快感原理を学んで自己成長に活かす(in Japanese)デゲーミフィケーションとは何か ゲームの快感原理を学んで自己成長に活かす(in Japanese)
デゲーミフィケーションとは何か ゲームの快感原理を学んで自己成長に活かす(in Japanese)
 
AWS greengrassとは何か (in Japanese)
AWS greengrassとは何か (in Japanese)AWS greengrassとは何か (in Japanese)
AWS greengrassとは何か (in Japanese)
 
シンギュラリティ哲学:超知能時代の教養の行方 (in Japanese)
 シンギュラリティ哲学:超知能時代の教養の行方 (in Japanese) シンギュラリティ哲学:超知能時代の教養の行方 (in Japanese)
シンギュラリティ哲学:超知能時代の教養の行方 (in Japanese)
 
APIエコノミーとは何か? それはどこへ続く道なのか(2017年) (in Japanese)
APIエコノミーとは何か? それはどこへ続く道なのか(2017年) (in Japanese)APIエコノミーとは何か? それはどこへ続く道なのか(2017年) (in Japanese)
APIエコノミーとは何か? それはどこへ続く道なのか(2017年) (in Japanese)
 
ゲーミフィケーションとは何か? (in Japanese)
ゲーミフィケーションとは何か? (in Japanese)ゲーミフィケーションとは何か? (in Japanese)
ゲーミフィケーションとは何か? (in Japanese)
 

Kürzlich hochgeladen

Call Now ≽ 9953056974 ≼🔝 Call Girls In New Ashok Nagar ≼🔝 Delhi door step de...
Call Now ≽ 9953056974 ≼🔝 Call Girls In New Ashok Nagar  ≼🔝 Delhi door step de...Call Now ≽ 9953056974 ≼🔝 Call Girls In New Ashok Nagar  ≼🔝 Delhi door step de...
Call Now ≽ 9953056974 ≼🔝 Call Girls In New Ashok Nagar ≼🔝 Delhi door step de...
9953056974 Low Rate Call Girls In Saket, Delhi NCR
 
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...
Dr.Costas Sachpazis
 
result management system report for college project
result management system report for college projectresult management system report for college project
result management system report for college project
Tonystark477637
 

Kürzlich hochgeladen (20)

NFPA 5000 2024 standard .
NFPA 5000 2024 standard                                  .NFPA 5000 2024 standard                                  .
NFPA 5000 2024 standard .
 
Coefficient of Thermal Expansion and their Importance.pptx
Coefficient of Thermal Expansion and their Importance.pptxCoefficient of Thermal Expansion and their Importance.pptx
Coefficient of Thermal Expansion and their Importance.pptx
 
Call Now ≽ 9953056974 ≼🔝 Call Girls In New Ashok Nagar ≼🔝 Delhi door step de...
Call Now ≽ 9953056974 ≼🔝 Call Girls In New Ashok Nagar  ≼🔝 Delhi door step de...Call Now ≽ 9953056974 ≼🔝 Call Girls In New Ashok Nagar  ≼🔝 Delhi door step de...
Call Now ≽ 9953056974 ≼🔝 Call Girls In New Ashok Nagar ≼🔝 Delhi door step de...
 
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptx
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptxBSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptx
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptx
 
VIP Model Call Girls Kothrud ( Pune ) Call ON 8005736733 Starting From 5K to ...
VIP Model Call Girls Kothrud ( Pune ) Call ON 8005736733 Starting From 5K to ...VIP Model Call Girls Kothrud ( Pune ) Call ON 8005736733 Starting From 5K to ...
VIP Model Call Girls Kothrud ( Pune ) Call ON 8005736733 Starting From 5K to ...
 
Thermal Engineering-R & A / C - unit - V
Thermal Engineering-R & A / C - unit - VThermal Engineering-R & A / C - unit - V
Thermal Engineering-R & A / C - unit - V
 
PVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELL
PVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELLPVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELL
PVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELL
 
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...
 
Java Programming :Event Handling(Types of Events)
Java Programming :Event Handling(Types of Events)Java Programming :Event Handling(Types of Events)
Java Programming :Event Handling(Types of Events)
 
Intze Overhead Water Tank Design by Working Stress - IS Method.pdf
Intze Overhead Water Tank  Design by Working Stress - IS Method.pdfIntze Overhead Water Tank  Design by Working Stress - IS Method.pdf
Intze Overhead Water Tank Design by Working Stress - IS Method.pdf
 
data_management_and _data_science_cheat_sheet.pdf
data_management_and _data_science_cheat_sheet.pdfdata_management_and _data_science_cheat_sheet.pdf
data_management_and _data_science_cheat_sheet.pdf
 
Online banking management system project.pdf
Online banking management system project.pdfOnline banking management system project.pdf
Online banking management system project.pdf
 
UNIT-II FMM-Flow Through Circular Conduits
UNIT-II FMM-Flow Through Circular ConduitsUNIT-II FMM-Flow Through Circular Conduits
UNIT-II FMM-Flow Through Circular Conduits
 
CCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete Record
CCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete RecordCCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete Record
CCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete Record
 
Thermal Engineering Unit - I & II . ppt
Thermal Engineering  Unit - I & II . pptThermal Engineering  Unit - I & II . ppt
Thermal Engineering Unit - I & II . ppt
 
Vivazz, Mieres Social Housing Design Spain
Vivazz, Mieres Social Housing Design SpainVivazz, Mieres Social Housing Design Spain
Vivazz, Mieres Social Housing Design Spain
 
result management system report for college project
result management system report for college projectresult management system report for college project
result management system report for college project
 
Glass Ceramics: Processing and Properties
Glass Ceramics: Processing and PropertiesGlass Ceramics: Processing and Properties
Glass Ceramics: Processing and Properties
 
Call for Papers - International Journal of Intelligent Systems and Applicatio...
Call for Papers - International Journal of Intelligent Systems and Applicatio...Call for Papers - International Journal of Intelligent Systems and Applicatio...
Call for Papers - International Journal of Intelligent Systems and Applicatio...
 
UNIT-IFLUID PROPERTIES & FLOW CHARACTERISTICS
UNIT-IFLUID PROPERTIES & FLOW CHARACTERISTICSUNIT-IFLUID PROPERTIES & FLOW CHARACTERISTICS
UNIT-IFLUID PROPERTIES & FLOW CHARACTERISTICS
 

An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordination

  • 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordination ISCIT-2018, Bangkok, Thailand Toshihiko Yamakami CTO Team, ACCESS Toshihiko.Yamakami@access-company.com 2018/09 Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 1 / 22
  • 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Outline Background Related Work Research Method Design Edge and Cloud Coordination Requirements Architecture Implementation Prototype Implementation Model Switch Performance Discussion Advantages of the Proposed Model Possible Mechanism for Stability Limitations Conclusion Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 2 / 22
  • 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Background Research Purpose The aim of this research is to develop device-edge-cloud coordination in edge AI chip-based intelligent IoT systems. Background Cloud-based AI has advantages of flexibility, capacity and throughput, but it has disadvantages of high communication costs and long delays. Edge represents a node which is connected to devices and a cloud. It has low latency direct connection to devices. It also has a connection to a cloud, but does not have the full capabilities of the cloud, It is promising to utilize an edge AI development environment where a framework of edge-cloud AI coordination is deployed. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 3 / 22
  • 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Related Studies Research on edge intelligent computing consists of three areas: a) feasibility and framework of edge computing: fine-grained edge offloading architecture [Cozzolino17], auction scheme in edge resource allocation [Prasad17], conditions for edge computing feasibility [Kalim17] b) AI in edge computing: latency-aware edge computing in video analytics[Yi17], distributed processing of a Deep CNN in smart camera [Castillo17], feasibility of AI and ML at edge computing [Sharma18] c) intelligent systems in edge-cloud coordination: healthcare optimization in wearable-edge-cloud coordination[Strässle17] docker overhead in Mobile Edge Computing [Avino17] The past research did not cover the special purpose AI-engine-chip-based device-edge-cloud coordination from the software perspective. The originality of this paper lies in its identification of software-based approach of AI-engine-chip-based device-edge-cloud coordination. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 4 / 22
  • 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Research Method Identifying Edge AI requirements, Designing a Prototype Implementation of Edge-Cloud Coordinated AI, Implementing a Prototype using FPGA (Field-Programmable Gate Array) Discussing Lessons Learned in Prototype Implementation Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 5 / 22
  • 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Challenges of AI-engine-based AIoT systems Aspect Issues AI-Engine Induction performance, Power-consumption-efficiency, Low-latency, Privacy of data, Cost Cloud Learning efficiency, Performance of learning models Edge-Cloud coordi- nation Alignment of edge processing and cloud processing, Controllability of device execution using knowledge in the cloud. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 6 / 22
  • 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Requirements of Edge AI Computing Item Description Low latency Intelligent device control requires low latency. It is dif- ficult at cloud-based AI systems. Low communication costs Massive IoT systems need low communication costs in order to prevent cost explosion. The communication costs to the cloud can be a prohibitive factor. Privacy and security In-edge AI computing does not expose data to be pro- cessed. It is difficult at cloud-based AI systems. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 7 / 22
  • 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Advantages of edge-cloud coordination Item Description Utilization of Data Analytics The edge-cloud coordination enables utilization of stored learned models at the cloud. Coordination of dis- tributed edges The edge-cloud coordination enables real-time context utilization with aggregated data from distributed edges in a real-time manner. Task migration The edge-cloud coordination enables task migration and resource optimization among multiple edges. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 8 / 22
  • 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Use cases of Edge-Cloud Coordinated AI Item Description Data sharing When necessary, data at the edge need sharing with the cloud. Cloud-based control When a context is recognized by a cloud, a cloud-based control needs coordination at the edge side so that an intelligent cloud-based control is facilitated. For exam- ple, learned models can be switched according to the request from the cloud. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 9 / 22
  • 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A prototype environment Item Description Simulating environ- ment Xilinks ep706-based DMP evaluation SDK kit. CPU FPGA embedded OS Xilinks Linux OS Simulated core DMP IP Core Languages C++ to control device logics, Python to interact with a cloud. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 10 / 22
  • 11. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Architecture Overview Device (Camera, sensors, …) Edge (AI engine) Cloud (storage, management) Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 11 / 22
  • 12. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Edge Architecture High-level Modules AI Module Induction Module Learned Model Manager Cloud Coordination Module Low-level Modules Communication Module Local Storage Module Camera Module Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 12 / 22
  • 13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Implementation Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 13 / 22
  • 14. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Demo features implemented on a prototype environment Item Description Video recording Video recording is performed with a camera connected to the device. Object detection Objects are detected by the pre-set learned models. The number of detected objects are counted and compared with the pre-set threshold. Image sharing Images with rectangle-framed detected objects are up- loaded. Remote setup of learned models The learned models can be switched by a command from a cloud using a switch-and-reset command. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 14 / 22
  • 15. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Model Switch Performance This paper does not cover detailed performance analysis of the current implementation. For model switch time, It takes approximately 15 second to switch learned models In the current implementation. The major time consumption lies in a) model load and execution and b) waiting time for graceful exit for model loading. Without the latter, the execution of new learned model is unstable. Also, the acknowledgement of model switch command takes 3 seconds at maximum. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 15 / 22
  • 16. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Advantages of the Proposed Approach The implementation of a prototype in FPGA-simulated IP core verifies points (see Result Table) The demo demonstrates the feasibility of edge-cloud coordination using ASIC-based AI engines as edge computing. There are lessons learned in the implementation on needs of software-based support for hardware-involved integration (see Lessons Table). Possible mechanisms to enhance stability of edge AI computing are needed (see Possible mechanisms table). Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 16 / 22
  • 17. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Results of the implementation Item Description Edge-closed process- ing Low latency, privacy, and communication-less process- ing are implemented. High-level learned model management It is feasible to switch multiple learned models depend- ing contexts. On-demand infor- mation sharing When a context requires full image transfer, needed images can be uploaded to the cloud. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 17 / 22
  • 18. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lessons learned in the implementation Item Description Timing to ASIC It is critical to align some hardware timing. When the timing is not aligned, a switch of learned models does not work and the system is frozen without any outputs. Timing coordination between an edge and a cloud In an object detection use case, time alignment between an edge and a cloud is subtle. The slight time differ- ence gives a small discrepancy between the detected rectangles and the transferred image. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 18 / 22
  • 19. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Possible mechanisms to enhance stability of edge AI computing Item Description Debug Logging, tracing, and probing functions are provided. Self-diagnosis Self diagnosis of monitoring process is performed in or- der to identify induction process faults without human intervention.. Adaptive computing Fault-tolerant processing when some parts of camera- AI process chain do not work properly. Remote setup of learned models The learned models can be switched by a command from a cloud using a switch-and-reset command. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 19 / 22
  • 20. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Limitations This implementation is a proof of concept implementation. Camera works and detected data can be shared, but performance, scalability, and reliability are out of the scope of this paper. Function is based on conceptual verification. Command protocol is not implemented and interaction with the cloud is performed by the state variable control at the cloud. Images are not encoded at the device and shared by a binary array format. In-depth synchronization of detected objects and shared images remains for future studies. This research is a prototype implementation based on a simulated environment constructed on ASIC. The real world deployment will be based on specialized battery-efficient Edge AI processor core. The system performance for such specialized AI hardware remains future studies. Theoretically, model conversions from Caffe should work, but it does not perform reliably on the implementation in this paper. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 20 / 22
  • 21. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion Requirements of low latency, privacy, and limited bandwidth drive the shift from cloud computing and edge computing. In this paper, the design of an experimental prototype for device-edge-cloud coordination for object detection use cases is presented. The feasibility of device-edge-cloud coordination is verified by an implementation using FPGA. It shows the effectiveness of edge-based real-time object detection with high-level cloud-base manipulation of multiple learned models stored in an edge node. This experimental implementation can lead to the full ASIC-based implementation when the IP core is ASIC-fabricated. At the same time, the experiment exposes challenges in hardware-software coordination. The author discusses design of software-based compensation towards these challenges. The design derived from this implementation is a stepping stone for full ASIC-based AI-engine solutions in the device-edge-cloud coordination. Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 21 / 22
  • 22. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Thank you for your attention! Questions? Toshihiko Yamakami (ACCESS Confidential)An Experimental Implementation of an Edge-based AI Engine with Edge-Cloud Coordinatio2018/09 22 / 22