la m and e symposium nov 2017 all decks.pdf
TRANSCRIPT
© 2017, Amazon Web Services, Inc. or its Affiliates.
November 7th 2017
AWS Media & EntertainmentLA Media Symposium - Loews Hollywood
© 2017, Amazon Web Services, Inc. or its Affiliates.
© 2017, Amazon Web Services, Inc. or its Affiliates.
Today’s Agenda: AM
Time General Track
10:00 AM Registration
10:30 AM Welcome - AWS Update
Bhavik Vyas, AWS Global Partner Segment Lead for M&E
10:45 AM Hulu Live TV on AWS
Rafael Soltanovich, VP of Software Development for Hulu
11:00 AM Broadcast Media Management in the Cloud
Chris Blandy, EVP Technology Solutions for Fox
11:30 AM Securing Content torage and Workflows to Meet Studio Requirements
Eli Mezie, Lead Architect Executive at ISE
Dan Meacham, CISO for Legendary Pictures
12:00 PM Networking Lunch
© 2017, Amazon Web Services, Inc. or its Affiliates.
Today’s Agenda: PMTime Track #1: Production & Content Creation Track #2: Broadcast & OTT Track
1:00 PM
Track Session #1
Rendering in Cloud with Deadline 10
Chris Bond, AWS Thinkbox CEO
Jason Fotter, Co-Founder and CTO, FuseFX
Track Session #1
AI Distribution - C-SPAN
Dan Mbanga, AWS BD Artificial Intelligence
1:45 PM
Track Session #2
Panel Discussion: Remote Creatives - Storytelling In a Distributed Age
Igor Boshoer, Founder/CEO for Linumio;
David Benson, CTO for Bebop
David Andrade, Cofounder for Theory Studios
Andy Stack, COO for Arcturus Studio
Track Session #2
Real-time Media Analysis with AI in Media
Kon Wilms, AWS Specialist Media Solutions Architect
2:30 PM Break (15mins)
2:45 PM
Track Session #3
Enhanced Content Workflows using Amazon AI
Kon Wilms, AWS Specialist Media Solutions Architect
Track Session #3
Delivery Side Supply Chain and OTT
Ken Kanagaraj, Principal Solutions Architect,
AWS Elemental
3:15 PM
Track Session #4
attn: Scaling to 500 million monthly views
Jake McGraw CTO Attn
Peter Agelasto, Founder Digital ReLab;
Luke Adams, CTO Digital ReLab;
Track Session #4
Low Latency Streaming via AWS and Wowza
Mike Talvensaari, Wowza VP of Product;
Chris Gruiz, WW Business Dev & Ops, AWS Marketplace
4:00 PM Closing Comments
4:15 PM Networking Reception
© 2017, Amazon Web Services, Inc. or its Affiliates.
L e a r n m o r e a b o u t t h e i n n o v a t i v e M & E s o l u t i o n s b u i l t o n A W S b y v i s i t i n g
o u r p a r t n e r s p o n s o r s i n t h e r e c e p t i o n a r e a d u r i n g t h e i n t e r m i s s i o n .
T H A N K Y O U
© 2017, Amazon Web Services, Inc. or its Affiliates.
Content Acquisition
Production/Post
Playout & Distribution
DAM/MAM/ Archive
OTT
Digital Supply Chain
Publishing
Analytics
AWS Enabled Media Workloads
© 2017, Amazon Web Services, Inc. or its Affiliates.
Ingest/Move Store Transform Process
LIVE OR FILE
WAN OR LAN
BLOCK OR OBJECT POST PRODUCTION
CONTENT FINISHING
LIVE OF FILE
PACKAGE & DELIVER
Basic Media Pipeline
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
Supply Chain Is the Media Dynamo
Playout &
DistributionLive TV
B2B/Cinematic
DeliveryDPP, IMF, NABA
Publishing
OTTLive-To-VOD
AVOD, SVOD, TVOD
Catchup TV
Cloud PVR
Normalization
Digital Supply Chain
Ingestion
Manual QC
Auto-QC
Metadata Extraction
Transcoding
Delivery
Packaging
Clips/Edits
CMS/DAM
Production/Post
ProductionEditorial
Dailies
Post
Social Media
Marketing/Promo
DAM & ArchiveActive Archive
Golden Copy
Preservation
Content AcquisitionLive
File
© 2017, Amazon Web Services, Inc. or its Affiliates.
Media Challenges and Cloud Benefits
Content Growth
More content, higher
quality (4K, HDR…).Scalable Storage
Peak DemandUnpredictable,
bursty workloadsElastic Compute, Pay
as you go
Competitive PressuresChanging consumption
patternsAgility and Innovation
Core CompetencyDatacenters vs.
core bizOutsource
© 2017, Amazon Web Services, Inc. or its Affiliates.
Approaches to Running Media Workloads on AWS
Custom APIs & code base
3rd party/ Open Source
BYO Interface(s)
CUSTOM:
BUILD YOUR OWN
Multiple
Inputs
AWS Elemental APIs
For Live, File, SSAI, OTT
3rd party
Svcs
Appliance VM
On-Premises
AWS ELEMENTAL
MEDIA SOLUTIONS
KEY AWS
PLATFORM
SERVICES
S3/
Glacier/
EFS
EC2Cloud
Front
RDS,
Aurora, Redshift,
DynamoDB
Lambda
SWF
AML/
KinesisRekognition
Polly/ Lex
Deadline
Pipeline Scheduler
Render/VFX Workloads
(Onprem, Hybrid, Cloud)
THINKBOX
Thinkbox MarketplacePAYG 3rd Party Software
BYOL, S/PaaS, MSP,
Media, Security,
Devops, Monitoring,
Data Analysis,
Automation
PARTNER-BASED
AWS MarketPlace
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
Common Digital Media Workflows on AWS
File transfer (DX, S3
transfer
accelerator,
Aspera,
Signiant…)
Ingest
Snowball,
Snowmobile
Live stream (e.g. AWS
Elemental)
Manage
Rights, Tag,
Admin
Analyze/ AI/ Monetize
Collect, Process, Analyze (Kinesis,
Redshift, AML, Rekognition…)
Amazon S3, IA, Glacier, EBS, EFS
Store
Process
AWS Elemental
Transform, QC,
Protect, Package
Stream (OTT),
Export, Playout,
Publish
Deliver
Edit,
Review,
Model…
Visualize
© 2017, Amazon Web Services, Inc. or its Affiliates.
Performance Efficiency: Go Serverless
AWS Lambda Amazon DynamoDBAmazon S3
Compute Storage Database
Amazon Kinesis
Orchestration & State Management Monitoring & Debugging
AWS X-RayAWS Step Functions
Analytics
Amazon SNS
Amazon API GatewayAmazon SQS
API Proxy Messaging & Queues AI/ML
Amazon Rekogniton
Amazon Polly
© 2017, Amazon Web Services, Inc. or its Affiliates.
AWS Well-Architected Best Practices
Reliability
Build for
failure
Performance
Efficiency
More with
less
Cost
Optimization
Margin &
Bottom-line
Operational
Excellence
Build, Run,
Monitor,
Scale
Security
Job #1.
© 2017, Amazon Web Services, Inc. or its Affiliates.
AWS Migration and Transition Methodology
HOW?
STAGES OF ADOPTION
What and Why?
AWS Cloud Adoption Framework (CAF) v2
Bu
sin
ess
Ca
pa
bil
ity
Fo
cuse
d
BusinessValue Realization
PeopleRoles and Readiness
GovernancePrioritization and Control
Te
chn
ica
l C
ap
ab
ilit
y
Fo
cuse
d
PlatformApplications and
Infrastructure
SecurityRisk and Compliance
OperationsHybrid and Dynamic
How and Who?
© 2017, Amazon Web Services, Inc. or its Affiliates.
NASA/JPL
AWS Media Customers
© 2017, Amazon Web Services, Inc. or its Affiliates.
Playout & Distribution
Discovery
Communications
Post Production
Scripps, Sony MCS,
FuzeFX, Bebop
Analytics/ AI
C-Span, MLB,
Channel 4
DAM & Archive
PBS, Charlie Rose
(Wazee), Discovery,
Sony DADC, Turner
Media Supply Chain
Turner (CNN), Sony DADC,
Discovery, Fox
Publishing
The Guardian,
Financial Times,
Gannett, Time Inc.,
Newscorp
OTT
Hulu,
Netflix,
MLB, BBC,
PBS
Acquisition
Scripps
AWS Customer Workload Adoption
© 2017, Amazon Web Services, Inc. or its Affiliates.
Hulu Live TV on AWS
Rafael Soltanovich, VP of Software Development
S o W h a t I s H u l u ?
A N e w H u l u E x p e r i e n c e W i t h
L i v e T V
We Solved A Ton Of Challenges In Less Than A Year
Greater than
1,000 streamsPersonalization and
recommendation
MPEG - DASH
for live
Commercial DRM
encryption
Metadata
reconciliation
H u l u ’ s F i r s t L a r g e -S c a l e D e p l o y m e n t
I n t o T h e C l o u d
H u l u L i v e O n A W S
Internet gateway
Multi-AZ
ALB
Amazon CloudFront(and others)
Manifest
RDS Aurora EC2
AWS
VPC
Amazon S3
Packaging
SegmentData
EC2
Live DB
ElastiCache
Users
Internet Gateway
Encoding
Vendors
ManifestPackaging
DirectConnect
RDS Aurora EC2
SegmentData
EC2
Live DB
ElastiCache
A W S P a r t n e r s h i p B e y o n d T h e T o o l s
Operational readiness
Performance tuning and optimization
A W S P a r t n e r s h i p B e y o n d T h e T o o l s
Launch support
TCO optimization
We’re Excited To Continue
Hulu Innovations In The Cloud!
© 2017, Amazon Web Services, Inc. or its Affiliates.
Broadcast Media Management in the Cloud
Chris Blandy, EVP Technology Solutions, Fox
Broadcast Media Management in the Cloud
Chris BlandyEVP, Technology Solutions
Fox Networks Engineering & Operations
Background on Fox NE&O
Our Mission: Technical operations platforms for Fox Networks Group
- Studio Production Facilities- Editing and Graphics Creation- Master Control / Playout- Linear Channel Distribution- Media Services- VOD
Our Technology Strategy
‘2020 Vision’
• Self-service tools upstream
• Assembly downstream
• Automation
• Cloud First
Why Cloud?
• Lower overhead
• Faster deployment
• Scale
• Security
Current Cloud Applications
Ingest and QC Automation
• By EOY 2017 ALL program and promo deliveries will flow through
AWS
• Delivery to our 2 primary playout centers in the formats we need
• Most assets pass QC checks without human intervention
• Instant replication = operational resiliency
Cloud Rendering
• Deadline Render Manager
• Enables VFX teams to use virtually unlimited render capacity
• Supports on-prem and cloud render farms
• True ‘burst’ capacity management
• Operations teams are in full control of where jobs run, how many instances, etc.
LTO Replacement
-
0.500
1.000
1.500
2.000
6/1/17 7/1/17 7/31/17 8/30/17 9/29/17 10/29/17
Storage Used (Petabytes)• We’ve achieved a truly ’tapeless’ environment
except for LTO tapes!• LTO very cost-effective for archive,
but over time we started using as a Tier 2 storage pool
• Instead of dozens of daily operations, we were seeing hundreds!
• Inevitable failures resulted• Solution: migrate working library to
S3/Glacier
OTT Distribution
• Dramatic change to Pay TV ecosystem• Initial approach included ‘TV Everywhere’
authentication• Digital MVPD launches:
• 2 in 2015• 1 in 2016• 3 in 2017
• Problem: most don’t have an easy way to acquire our network affiliate signals
OTT Distribution
Future Cloud Applications
Automated Versioning & Packaging
• Our assets are increasingly already in the cloud
• Vendor systems for media processing are migrating
• Transcoding options are mature
• Next steps:• Work order management
• Metadata drives versioning
• Object-based container format like IMF
Non-linear Editing
• Why not enable editing anywhere
• Current workflows require pushing media back and forth between edit workstations and production or playout
• Some ‘cloud’ options – but often proprietary
• Testing zeroclient with AWS WorkSpaces for NLE workflows
• Assets and metadata live in AWS, edit anywhere, full control
Machine Learning / AI
• Limited metadata exists if we have it at all!
• Can we apply ML/AI to generate richer metadata?
• AWS Rekognition• Frame-level image analysis
• Auto-tag scenes with talent metadata (RecognizeCelebrities API)
• Other applications:• Find efficiencies by mining workflow analytics
• Automate logging
• Predictive archive retrieval based on news headlines
• Social media sentiment analysis
Cloud Playout
• Final step to full cloud migration
• Recipe: Program assets in S3
Promos in S3
Ad creatives in S3
Log delivery to cloud
Playout automation
Encoding
Secure, error-free delivery to partners
• Gap remains – Live feed integration
Thank you!
© 2017, Amazon Web Services, Inc. or its Affiliates.
Securing Content Storage and Workflows to
Meet Studio Requirements
Eli Mezie, Lead Architect Executive at ISE
Dan Meacham, CISO for Legendary Pictures
S e c u r i n g c o n t e n t s t o r a g e a n d w o r k f l o w s t o m e e t s t u d i o r e q u i r e m e n t s
Dan Meachem
@DanTechieTweets
Eli Mezei
@ISESecurity
Introductions
46
Dan Meacham
CISO– Legendary Pictures
Dan oversees cyber security and content protection for Legendary, a
nearly 100% cloud-based organization.
Eli Mezei
Executive Partner – Independent Security Evaluators
Eli Mezei develops and executes advanced security programs
for clients across several industries, including especially media &
entertainment.
Workflows Dr ive Secur ity
• Security must support the
workflow, not the other way
around
• The workflow must be understood
in depth before security controls
can be defined
• The simplest solution is generally
the most secure
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Trust Model Vs. Threat Model
Know Your Adversary
On-Location Production Facility
On site Networked Digital Storage and Collaborative Processing Environments
Off-Site
On-Location Film Set
Audio Storage
Compact FlashSound
Film Set
Actor
AVID/ISIS
Vendors
VFX, Graphics,
Arts, Set Design,
etc.
Long Term
Offsite Storage
Second Unit
Production
HQ Production
and Reviews
Actor
On Set Review
Video Assist er
Qtake Client
Director
Qtake Client
Asst. Directors
Qtake Client
On-Set Crew On-Set Crew
Color Corrected
Live feed - SDICameras
Raw video via SDI
Raw Camera Capture
SSD Mags to
Production Facility
Network Appliance -
On-Set Disaster Recovery
DIT
Color Correction
on Set
Camera SSD
Misc. On-Set Digital Content
Creation and Reviews
Makeup Set Design Sound
Camera SSD
On Set VFX-Editorial Review
Director
VFX-Editorial Dailies
Asst. Directors Asst. Directors On-Set Crew On-Set CrewDOP
Camera SSD Mags
Editorial
Asst. Directors
Content Review
Content Editing
Editor
Dailies On-Set Review
VFX Editorial
VFX Editing VFX Review
VFX Editors
VFX and Art
Film VFX and Art
Set Design
VFX Art Set Designer
Production Operations
Script Distribution
Production Office Management
Contracts
On site Networked Digital Storage
File Server
Video Lab
Color Correction Verification
Video Lab Engineer
Archival (LTO) Database
In-house Film Lab
In house Post Production - Editorial
Ingest of Data to ISIS bucket from Camera storage
Transcode the Processed Film to formats and upload to ISIS
Post Production
Air-gapped Network
ISIS ISIS
Edit Scenes
Cut in VFX, Art, and Second Unit Footage
Prepare Review Cut
Post Production
Transcription
Sound Mixing
Color and Video Finishing
Translation
Graphics
Commercial Cloud ServicesPIX/DAX Filet rack
Speeder
bikeAspera
5th kind
Commercial
Simpl if ied Product ion Workflow1. Content Creation and Capture
2. On-Location Production –
Uses Cloud Based tools3. Post Production –
Relies Heavily on Cloud based Collaboration Tools
Zooming in a sub workflow – VFX
Off-Site
Editorial-VFX
Editing in ART
Editing in VFX
VFX
Post Production Departments
Editorial
Isolated NetworkISIS ISIS
Mount and Move of Original Video to staging area
Prepare Editorial Tools for Editing
Edit Film Scenes
Edit Film Sequences
Post Production
Off-Site Crew
Studio HQVFX Vendor
Art
WANCommercial Cloud
Collaboration Services
PixsFotokem ShotgunSpeederbike
Post Production/On-Location VFX-Editorial
Dailies Review Pre-visualization VFX ReviewEditorial Review
Physical Devices
On-Location Post Production Facility
EditorialArt Editorial-VFXVFX Vendor
Commercial Cloud Collaboration ServicesPixsFotokem Shotgun
Mac Mini Server
On-Premises Render Workf low
The Chal lenge
What are we trying to solve?
SPEEDStorage
S P A C E
Security
Render Work f low in AWS
Private Compute and Storage Subnet 10.0.2.0/24
AWS Batch
Spot Fleet
Spot Instances
instances
VPN Subnet 10.0.1.0/24
router
Bastion ServerAmazon CloudWatch
AWS KMS
Content Ingress/
Egress
Amazon EFS
route table
On-premises network
192.168.0.0/16
customer gateway
Active
Directory
Region
Amazon VPC
VPN gateway
VPN connection
route table
Compute Subnet
Render Farm Compute
Domain
ControllerAD
Connect
Sync
Access Control
Principals of Secure Design in Content
Product ion workf lows
Principal : Least Pr iv i lege
Principal : Pr iv i lege Separat ion
Robust Ident ity Controls
Subnet
AWS KMS
Region
IAM
Amazon VPC
route table
IAM
S3 Bucket
role
permissions
Users
Define role
Userpolicy
permissions
permissions
Encryptionkeys
AWS CloudFormation
Access Data
Monitor
Attach policy
AWS Encrypt ion & Key Management
AWS KMS
PRINCIPAL: DEFENSE IN DEPTH
VPC Segmentat ion
• Controls for an application or
publicly accessible service
• Controls that are currently used
can be applied to most AWS
implementations
VPC Peer ing
• Extends trust and controls
to Counterparties
• Restricts access to only
known infrastructure
• Used for workflows that
involve trusted counterparties
& shared infrastructure
VPC Sub-Account
• Trust is completely centralized
• Tantamount to a physical network
• Most restrictive trust model
Principal : Trust Reluctance
Principal : Trust Reluctance
Virtual private cloud
Amazon VPC
flow logsAmazon CloudWatch
bucket
AWS CloudTrail
Monitor API
VPC Logs
Asset Subnet
AWS KMS
route table
S3 Bucket
Encryptionkeys
Logging data
Key Logging
Logging datainstance with
CloudWatch
Amazon
Elasticsearch
Service
Event drive analytics
analytics'
Amazon RedshiftData warehousequeries
AWS LambdaCentralized Logging Proxy
Proxy ServerAdministrator
Log Management
API Calls
Summing it a l l up
SECURE DESIGN IS INTERCHANGEABLE
What’s next
• Collaborate with vendors and with productions to identify workflows to port to
the cloud
• Use published guidance
• Leverage AWS services and documentation
• Blaze a new trail
© 2017, Amazon Web Services, Inc. or its Affiliates.
Afternoon AgendaTime Track #1: Production & Content Creation Track #2: Broadcast & OTT Track
1:00 PM
Track Session #1
Rendering in Cloud with Deadline 10
Chris Bond, AWS Thinkbox CEO
Jason Fotter, Co-Founder and CTO, FuseFX
Track Session #1
AI Distribution - C-SPAN
Dan Mbanga, AWS BD Artificial Intelligence
1:45 PM
Track Session #2
Panel Discussion: Remote Creatives - Storytelling In a Distributed Age
Igor Boshoer, Founder/CEO for Linumio;
David Benson, CTO for Bebop
David Andrade, Cofounder for Theory Studios
Andy Stack, COO for Arcturus Studio
Track Session #2
Real-time Media Analysis with AI in Media
Kon Wilms, AWS Specialist Media Solutions Architect
2:30 PM Break (15mins)
2:45 PM
Track Session #3
Enhanced Content Workflows using Amazon AI
Kon Wilms, AWS Specialist Media Solutions Architect
Track Session #3
Delivery Side Supply Chain and OTT
Ken Kanagaraj, Principal Solutions Architect,
AWS Elemental
3:15 PM
Track Session #4
attn: Scaling to 500 million monthly views
Jake McGraw CTO Attn
Peter Agelasto, Founder Digital ReLab;
Luke Adams, CTO Digital ReLab;
Track Session #4
Low Latency Streaming via AWS and Wowza
Mike Talvensaari, Wowza VP of Product;
Chris Gruiz, WW Business Dev & Ops, AWS Marketplace
4:00 PM Closing Comments
4:15 PM Networking Reception
© 2017, Amazon Web Services, Inc. or its Affiliates.
L e a r n m o r e a b o u t t h e i n n o v a t i v e M & E s o l u t i o n s b u i l t o n A W S b y v i s i t i n g
o u r p a r t n e r s p o n s o r s i n t h e r e c e p t i o n a r e a d u r i n g t h e i n t e r m i s s i o n .
L U N C H : R E S T A R T @ 1 P M
Low-Latency Streamingvia AWS and Wowza
Founded in 2005, Wowza offers a complete portfolio to power today’s video streaming ecosystem from encoding to delivery. Wowza provides both software and managed streaming services for producers, developers and engineers to build unique streaming experiences for any device or audience size. Wowza Streaming Engine is Wowza’sflagship award-winning media server software. Wowza Streaming Cloud is an end-to-end live streaming service built on AWS. Wowzaalso offers Wowza GoCoder a mobile SDK and free app for live streaming and Wowza Player, a modern HTML5 video player that is tightly integrated with other Wowza products.
What is Latency?
• Latency is a time interval between the stimulation and response, or, from a more general point of view, a time delay between the cause and the effect of some physical change in the system being observed.
• In terms of streaming, latency is the delay between the initial capture of the video and the viewer.
Other Related Terms
• Time to first frame - Time delay from when a person clicks the play button and when video appears
• Broadcast Delay – The practice of intentionally delaying the broadcast of live material to prevent profanity, bloopers, or violence (wikipedia)
• Drift – when latency increases during a broadcast
• Quality - Higher quality = higher resolution = more data to send
• Scale - How many inputs/participants, how far, how many viewers
• Bandwidth - How much traffic can the infrastructure ideally handle
• Throughput - How much traffic is the infrastructure really delivering
• Bitrate - How many bits of data are being processed over time
Why is there latency?
Q: Why is latency a problem?
A: Latency is not usually a problem – but people think it is.
• For many (or possibly most) live streams, latency does not matter
• HTTP streaming intentionally introduces latency for improved reliability.
• For some live streams, latency is critical.
Latency Demo
What Factors Impact Latency
Main factors affecting latency are:1. Quality2. Scale
Low Latency
Large Scale
High Quality
As either (or both) increase latency increases
What Factors Impact LatencyLow
Latency
Large Scale
High Quality
Quality• Resolution• Two-way• Multi-thread• Frame rate• Smooth playback• Chunk/Block Size• Buffering
Scale• Distance• Participants• Viewers• Streams• Complexity• Variability• Diverse endpoints
How Quality Injects LatencyHigher quality = higher resolution = more data to send• Resolution increases the amount of data in any frame, block, chunk or time
slice. If we hold infrastructure constant as resolution goes up the window and latency to deliver a segment increases.
• Trade-offs: Some mitigation techniques
Increasing ThroughputIncreasing throughput with more bandwidth(network) and bitrate(compute) = More infrastructure, more overhead, more cost
Reducing BufferReducing the buffer requirement will reduce latency but it also makes network fluctuations more visible to the viewer – e.g. a 1ms buffer ideally means content can be viewed as little as 1ms after capture but a 0.1-second network interruption will result in 100 buffer segments being un-recoverable in a live stream.
Reducing FramerateReducing the (frame rate) number of frames per second reduces the smoothness of the viewing experience – e.g. 15 fps is generally a good viewing experience for low movement broadcasts like presentation screen capture/sharing or many church services, it will produce a choppy experience for viewing sports
How Scale Injects LatencyScale increases distances, complexity and variability• As distance, the number of streams, and viewers increase, network
imperfections, variability, and degradation increase and can amplify latency.
• Trade-offs: Some mitigation techniques
Increasing ThroughputIncreasing throughput with more bandwidth(network) and bitrate(compute) = More infrastructure, more complexity, more overhead, more cost
Increasing BufferIncreasing buffer size adds ability for both encoders, transcoders and players to deal with variability smoothly but increasing buffer inherently increases latency. By definition, buffer size is the minimum amount of data for which a process can be initiated. If more data must be accumulated, then processing will take longer.
TCP vs UDPBecause it is a lighter weight protocol without error checking, monitoring, order of messaging and headers roughly 1/3 the size of TCP, UDP is inherently faster but there is no guarantee data is received. SRT can help here
Where streaming latency has been
• Windows Media• MMS, RTSP, HTTP (using UDP or
TCP)
• Encoder, Server, and Player buffer size management
• “Low delay” audio codecs
• Fast Start, Advanced Fast Start, Fast Recovery, Fast Reconnect, FEC
• 2-3 seconds (on a good network)
• Real Time Streaming Protocol (RTSP)
• Developed and by RealNetworks
• Relies on RTP and RTCP
• Frames can be sent one at a time in real time
• Can leverage UDP
• Potential latency could be as low as ~125 ms with minimal buffering (2-3 frames behind)
Where streaming latency has been
• Real Time Messaging Protocol (RTMP)
• Developed and open sourced by Adobe
• 1-3 seconds and sometimes below 1 second
• HTTP• Apple HLS
• 30+ seconds on iOS (using 10 second chunks)
• MPEG-DASH• 10-20 seconds (variable)
Where latency is going
• WebRTC• Designed for real-time audio,
video and data delivery over less-reliable connections
• Leverages TCP or UDP
• Multiple protocols related to RTSP/RTP
• 1 second or less (as low as 200 ms)
• WebSocket• Designed to provide a
standardized, two-way, reliable communications channel between a browser and a server
• Works with TCP
• Can be used with other streaming protocols including RTMP, WebRTC, Haivision SRT, Wowza WOWZ and Aspera FASP
• As low as 200 ms
Where latency is going
• HTTP• Reducing to 4 seconds (using 1 second chunks) for DASH and below 8
seconds (using 2 second chunks) for HLS (includes GOP adjustments as well)
• CMAF + HTTP/1.1 (using HTTP Chunked Transfer Coding) enables video transfer, decode, and display before the end of the chunk encoding (“chunks of chunks”)
• Optimizing DASH MPD attributes (availabilityStartTime and minBufferTime)
• Video encoding enhancements that do not impact decoding (H.264 GDR)
Where latency is going
• Quick UDP Internet Connections (QUIC)• Uses UDP (with TCP fallback)
• Focuses on security and reliability
• Tries to be like TCP while reducing connect and round trip times, packet loss, congestion control, and more using intelligent retransmissions and storage and delivery of information
• Built with HTTP/2 in mind
• And….
© 2017 Wowza Media Systems, LLC. All rights reserved. Confidential & Proprietary
Network Characteristic that KILL real time video
Packet loss – packets being discarded by routers
Jitter – packets arriving at different times than expected
Delay – the time from sender to receiver
Bandwidth – the fluctuating capacity between points
Overcoming Adverse Network Conditions
© 2017 Wowza Media Systems, LLC. All rights reserved. Confidential & Proprietary
SRT Fundamentals
Public InternetSource Destination
SRT Workflow
AWS AWS
© 2017 Wowza Media Systems, LLC. All rights reserved. Confidential & Proprietary
SRT Fundamentals
bps
tConstant frame rate
Encoder Variable Bitrate
Unreliable “Dirty" NetworkPacket Loss, Network Jitter,
Variable Delay, Bandwidth Fluctuation
tErratic bitrate and frame rate
Decoder’s Worst Nightmare!
SRT Fundamentals
© 2017 Wowza Media Systems, LLC. All rights reserved. Confidential & Proprietary
bps
tConstant frame rate
Encoder Variable Bitrate
tReconstructed video frame rate
Much Happier Decoder
SenderBuffer
ReceiverBuffer
SRTEncapsulation & Network Time Stamps
Feedback
SRT Fundamentals
SRT Fundamentals
© 2017 Wowza Media Systems, LLC. All rights reserved. Confidential & Proprietary
bps
tConstant frame rate
Encoder Variable Bitrate
tReconstructed video frame rate
Much Happier Decoder Video
Audio
Metadata
Packet acknowledgement – error recovery
Timing – low latency streaming
Network connection characteristics
Bi-directional
Firewall Friendly
SRT Fundamentals
SRT Fundamentals
SRT Demo
SRT in Wowza Streaming Engine
Greater reliability of throughput and packet loss
Improved bandwidth utilization with network bandwidth fluctuations & long-haul video transport
RESULTS - Germany to Australia, using a consumer DSL connection: • 7.5x increase of throughput with SRT, compared to RTMP connections over TCP.• RTMP: 800/10,000 kbits throughput• SRT: and 6000/10,000 kbits throughput
Independent security / encryption
Minimized buffer 4x RTT (re-transmission time)
Logic for packet loss recovery and jitter only requires 4 x the network latency to the server
Wide adoption: free open source, known & trusted
SRT in Wowza Streaming Engine Wowza Streaming EngineSRT Fundamentals
Benefits of SRT | Results
AND
WITH
AND
» Needs / Challenges
» Provide ESPN-like sports broadcasting to any screen
» High reliability and scalability to support load fluctuations
» Affordable for Division II & III universities and high schools
» Reliable first-mile network connections from remote venues
» Solution
» Wowza + AWS + SRT
» SRT integration significantly increases first-mile throughput
» Wowza transcoding enables ABR, simplifies on-site capture
» EC2 gives on-demand scalability to match broadcast load
» CloudFront supplies cost-effective global delivery
» Benefits
» Increases reliability of connections, reducing support calls
» Enables new opportunities for end users at remote venues
» Reduces costs from alternative solutions by 50%
Business Case
Enabling Customer Success for BlueFrame Technology
Common BlueFrame Technology Customer Scenario
Wireless connection without SRT HLS
SDI, HDMI, or NDI
EC2
S3
CloudFront
Wireless connection using SRT
Live broadcast from a college athletics venue with unreliable Internet connectivity
SRT Alliance MembersMembers
Try SRT in Wowza Streaming Engine
• Sign up for a free trial
• www.wowza.com/products/streaming-engine
• Images available in AWS Marketplace
Questions?
Thank you!
© 2017, Amazon Web Services, Inc. or its Affiliates.
AWS Media & Entertainment Cloud SymposiumKen Kanagaraj, Principal Solutions Architect, AWS Elemental
MEDIA SUPPLY CHAIN ON AWS
March 14, 2018
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
108
AWS acquired Elemental Technologies October, 2015
Enterprise Government
Broadcasters
Content Programmers
Pay TV Operators
MARKETSCONTENT
ON DEMAND
Storage
Amazon S3
LIVE
DISPLAYS
DEVICES
AWS ELEMENTAL – PERFECTING THE MEDIA EXPERIENCE
© 2017, Amazon Web Services, Inc. or its Affiliates.
THE ECOSYSTEM: POWERING MEDIA WORKFLOWS
Content
Production
Content Ingest
& Storage
Processing
&
Management
Content
Distribution &
Consumption
Consumer
Insight &
Analytics
Partner Solutions
Shared Services
Infrastructure Security Network Operations
• Modelling
• Rendering
• Video Editing
• Post Production
• Broadcast Signal
Acquisition
• Digital Dailies /
Approvals
• High Speed Ingest
• Library Storage &
Archive
• Tier Management
• Content / Asset
Management
• Machine Learning /
Artificial Intelligence
• En/Transcode
• Caption, Subtitles
• Packaging
• Encryption;
Watermarking
• Digital Rights
Management
• Automation,
Workflow
• Machine Learning /
Artificial Intelligence
• B2C Streaming of
Live and VOD
Content
• B2B Distribution
• Video Advertising
Insertion
• Analytics, Reporting,
Log Analysis
• Real-time Monitoring
• Content Discovery
• Content
Recommendations
• Machine Learning /
Artificial Intelligence
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
DELIVERY MEDIA SUPPLY CHAIN
Ingestion
Manual QC
Media Info
Metadata Extraction
Transcoding
Delivery
Packaging
Cataloging
CMS/DAM
Normalization
Content Acquisition
Post Production
DAM & Archiving
Playout & Distribution
OTT
Analytics
Media Supply Chain
“The digital supply chain is new media term which encompasses the processes involved for
the preparation and delivery of media, be it video or music, by electronic means from point of
origin (content provider) to destination ( consumer) “
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
111
Content&AssetManagement
ContentPrep&DeliverySources
Entitlements
Multi-DRM
SubscriberManagement
DeviceManagement
MetricsandAnalytics
Search&Discovery
Events/AlertsHeartbeat
ProductCatalogs
Authentication
WebCMS/AppPublishing
Recording/Scheduling
CustomerOSS/BSS Recommendations GeoServices SocialNetworks
Self-Care
AdServerMetadataEnrichment
VideoBackOffice AggregationAPIs
Personalization
OffersandPackages
DVRSchedulingRecordings
Tablets
Mobile
SmartTVs
IPSTB
GamingConsoles
Client/PlayerApplications
PC/Mac
Customerand3rd PartyInterfaces
JITP
DAI/StreamConditioning
IngestWorkflows
AWSConsoleAPIs
Logging OperationalConsoles NetworkMonitoring VideoQuality
VideoApplications,NetworkandBroadcastOperations
CDN
LiveTranscoding
VODEncoding
JSONAPI
ADIXMLTV
APIJSON/XML
AWSCoreServices
LiveFeeds
VODAssets
ADI/EPGMetadata
ContentPolicies
EC2 S3 DynamoDB CloudWatch Lambda
Tagging/Categories/Discovery
AnalystApplications
E2EOTTArchitecture
Billing
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
10
CASE STUDY: NEWS BROADCAST PROVIDER
Background:
Existing on-premises workflow to enable news producers from
all over the globe to submit videos to be processed and prepped
for OTT delivery across multiple devices.
Challenge:
Outgrown capacity on current infrastructure. Processing 200-
300 clips / day. Improve ‘time-to-web’.
Solution:
Develop content workflow leveraging Elemental Cloud, AWS native
components along with 3rd party offerings to increase capacity, handle
unpredictable workloads and meet SLA requirements – all by paying for
what they use.
© 2017, Amazon Web Services, Inc. or its Affiliates.
MEDIA SUPPLY CHAIN
Ingest QCInput AnalysisOutput
ProcessingDeliver
WORKFLOW STATE MANAGEMENT
Asset Registration
CMS
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
• Build it yourself vs. PaaS vs. SaaS
• Existing resource skills
• Elemental Cloud vs. partner/competitor processing
• Redundancy models
• No Multicast in the Cloud!
DESIGN CONSIDERATIONS
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
More than 90 services from
• compute,
• storage,
• networking,
• database,
• analytics,
• application services,
• deployment, management,
• Internet of Things (IoT),
• Artificial Intelligence (AI),
• security, hybrid and enterprise applications
SERVICE BREADTH
© 2017, Amazon Web Services, Inc. or its Affiliates.
• Dedicated, 1 or 10 GE private pipes into AWS
• Create private (VPC) or public virtual interfaces to AWS
• Reduced data-out rates (data-in still free)
• Consistent network performance
• At least 1 location to each AWS region
• Option for redundant connections
• Uses BGP to exchange routing information over a VLAN
INGEST
WHAT IS AWS DIRECT CONNECT?
© 2017, Amazon Web Services, Inc. or its Affiliates.
S3 Standard S3 Standard - Infrequent
AccessAmazon Glacier
Data Transfer (Ingest/Egress)
AWS Direct
Connect
AWS Snowball ISV Connectors Amazon Kinesis
Firehose
S3 VPC
EndPoint
AWS Storage
Gateway
Events
S3 Event
Notifications
S3 Transfer
Acceleration
STORAGE
OBJECT STORAGE OPTIONS
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
MEDIA WORKLOADS (REDEFINED)
EBS
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
119
AWS ELEMENTAL ENABLES ELASTIC, FLEXIBLE, MODULAR WORKFLOWS
PROCESS
AWS Elemental LIVEReal-time video processing
for linear broadcast
AWS Elemental DELTAOrigin and video delivery platform
AWS Elemental CLOUDManaged service providing video
processing and delivery in the cloud
AWS Elemental SERVERFile-based video processing for broadcast
and IP content
AWS Elemental CONDUCTORUnified video management
DELIVERYPROCESSING AND MANAGEMENT
INTEROPERABLE COMPONENTS
SUPPORTED INFRASTRUCTURES
Ad Decision CDN CMS/WFM DRM Fingerprinting QC Storage Watermarking
DEVICESDEVICES
AWS Elemental SERVER
AWS Elemental LIVE
AWS Elemental CONDUCTOR LIVE
AWS Elemental CONDUCTOR FILE
AWS Elemental DELTA
AWS Elemental DELTA
MANAGED NETWORK
AmazonCLOUDFRONT
SOURCE
FILE
LIVE
MPEG-2, AVC, HEVC, J2K
Storage
Amazon S3
Appliance On-premises
Virtual Machine AWS Cloud
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates. As of February 2017
DELIVERY
Ashburn, VA (3)
Atlanta, GA (2)
Chicago, IL
Dallas/Fort Worth, TX (2)
Hayward, CA
Jacksonville, FL
Los Angeles, CA (2)
Miami, FL
Minneapolis, MN
Montreal, QC
Newark, NJ
New York, NY (3)
Palo Alto, CA
San Jose, CA
Seattle, WA
South Bend, IN
St. Louis, MO
Toronto, ON
Rio de Janeiro, Brazil (2)
São Paulo, Brazil
Amsterdam, The Netherlands
(2)
Berlin, Germany
Dublin, Ireland
Frankfurt, Germany (5)
London, England (4)
Madrid, Spain
Marseille, France
Milan, Italy
Munich, Germany
Paris, France (2)
Stockholm, Sweden
Vienna, Austria
Warsaw, Poland
Chennai, India
Hong Kong, China (3)
Manila, the Philippines
Melbourne, Australia
Mumbai, India (2)
New Delhi, India
Osaka, Japan
Seoul, Korea (3)
Singapore (2)
Sydney, Australia
Taipei, Taiwan
Tokyo, Japan (3)
70 CloudFront Edge Locations (PoPs), 45 Cities, 5 Continents
CLOUDFRONT IS GLOBALLY DISTRIBUTED
© 2017, Amazon Web Services, Inc. or its Affiliates.
to manageSimple
Agile
Available
Scalable
Zero effort to administer
on demand
without ”clustering”
coding practices
Instance-lessCOMPUTE
© 2017, Amazon Web Services, Inc. or its Affiliates.
SERVERLESS – DELIVER ON DEMAND, NEVER PAY FOR IDLE
© 2017, Amazon Web Services, Inc. or its Affiliates.
BUILDING BLOCKS FOR SERVERLESS APPLICATIONS
AWS Lambda Amazon DynamoDB
Amazon SNS
Amazon API GatewayAmazon SQS
Amazon Kinesis
Amazon S3
Orchestration and State Management
API Proxy Messaging and Queues Analytics
Logging
Compute Storage Database
AWS Cloud
WatchAWS Step Functions
Processing
AWS Elemental
FfMEG
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
Amazon
STORAGE
Amazon
DELIVERY
Amazon
COMPUTE
Amazon
PROCESSING
Amazon
INGEST
CloudFront – Global Content Delivery Network with Monitoring, Analytics and customization at the edge
S3, Glacier, EFS & EBS – Durable, performant, scalable and secure solutions for on-line and archival content storage
Elastic Transcoder; AWS Elemental – Scalable, and cost-effective video processing on prem and in the cloud
EC2; Lambda– Resizable general purpose compute capacity featuring instance types optimized for rendering, editing, processing video, analytics
Direct Connect; Snowball; S3 Transfer Accelerator – Upload options for content and files of all sizes
Amazon
ANALYTICS/ AI
Machine Learning, Kinesis, AI – Data ingest/processing/analysis; machine learning, artificial intelligence for text to speech, image processing & analytics, natural language.
AWS SERVICES FOR M&E HIGHLIGHTS
© 2017, Amazon Web Services, Inc. or its Affiliates.
MEDIA ECOSYSTEM OF PARTNERS
INGEST STORE MANAGE SECUREPROCESS
CREATEMONETIZE
INTEGRATEDELIVER
SaaS BYOLPaaS
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
Source
Amazon S3
Media FileRegister
Asset
Media Info
QC
Encode
Publish
Elemental
Cloud
AWS Elastic
Transcoder
Output
Amazon S3
CloudFront
Amazon S3
• Images• Captions• Feature
Asset• Metadata
• MP4 –Bitrates
• ABR - HLS
Register
Asset
MediaInf
o
AWS
Step Functions
DynamoDB
Encode
HLS MP3
Publish
Q
C
VOD WORKFLOW – REFERENCE SOLUTION
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
AWS LIVE STREAMING
AWS provides a live streaming video solution that combines AWS Elemental Cloud, a service that enables customers to rapidly deploy multiscreen offerings for live and on-demand content, with Elastic Load Balancing, Auto Scaling, and Amazon CloudFront to build a highly resilient and scalable architecture that delivers your live
content worldwide.
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
• Modular
• Flexible – Avoid Vendor Lock In
• Elasticity
• Increased Development Velocity
• Leverage existing skills ( Java/Javascript)
NEW SOLUTION REQUIREMENTS
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
VOD DIGITAL SUPPLY CHAIN
Source
Amazon S3
Amazon
Dynamo DB
Amazon
CloudWatch
Dashboard &
ConsoleAmazon
SNS/SQS
Amazon
Step FunctionsAmazon SQS
Amazon Lambda
MediaInfo QC Integration
AWS Elemental xCode
AWS Elemental JITP ( Delta )
Outputs
Amazon S3
MP4 / HLS Output
Events / Logs Notifications
Multi-CDN
DRM
Service
Fast File
Transfer
DRM
Amazon
Elastic Search
Metadata Index
External Head End
Asset Metadata
Technical Metadata
Encoding Profiles
QC reports
© 2017, Amazon Web Services, Inc. or its Affiliates.
CLOUD ADOPTION APPROACHES
• Deploy application on cloud infrastructure services
Lift-and-Shift
• Refactor application to leverage some Cloud abstractions
Cloud-Optimized
• Re-imagine application architecture with many cloud abstractions
Cloud-Native
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
INGEST FLOW
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
PROCESS FLOW
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
PUBLISH
© 2017, Amazon Web Services, Inc. or its Affiliates.
BENEFITS AND CHALLENGES SOLVED
AWS helps you scale your media
storage and compute needs
Handle unpredictable &
bursty media needsPay for media you store
and process, as you go
Global availability instantly,
with no commit
Focus your resources on
your media needs
Shorten time-to-air,
increased agility, automate
Cost PressuresContent Growth Variable Demand
Operational Pressures Global Market Core Competency
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
AWS alignment to MPAA Security Best Practices reviewed October 2012, Re-Reviewed August
2015 Based on the MPAA's revised Best practices for
Content security and AWS Shared Responsibility Model
AWS Services In Scope
“All major AWS Services”
Content Types In Scope
RAW master, high/low-resolution, watermarked/spoiled, full/partial feature content, stills, clips, frames,
shots, sequences, scripts, storyboards, as well as production and post-production deliverable formats in
pre-and post-theatrical release windows
MPAA Hub page on AWS Compliance http://aws.amazon.com/compliance/mpaa/
SECURITY BEST PRACTICES – MPAA ETC.
© 2017, Amazon Web Services, Inc. or its Affiliates.
DYNAMIC/JUST-IN-TIME PACKAGING
Encoder Packager Origin Player
1 2
Pre-packaging
Costs +: Each segmented format stored separately. Store whole library 3-4 times.
Costs - : Stored Once, processed on demand.
3
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
MULTI-CDN – DYNAMIC PACKAGING
240 @ 600kbps
1080 @ 5Mbps
720 @ 3Mbps
576 @ 1.4Mbps
360 @ 1Mbps
AAC Stereo
Dolby Digital
STB & CONNECTED TV
VIDEO DELIVERY PLATFORM
1080 @ 5Mbps
720 @ 3Mbps
576 @ 1.4Mbps
360 @ 1Mbps
AAC Stereo
720 @ 3Mbps
576 @ 1.4Mbps
360 @ 1Mbps
240 @ 600kbps
AAC Stereo
There is no point in streaming HD
1080p content to a smartphone:
even if the device can render it,
users do not see the difference
with 720p on small screens.
Reduction in bandwidth costs
and initial video starting time.
On Delta, the same URL
http://<delta>/out/u/hls.m3u8
will automatically apply the
selection of the audio/video
streams, size of the video
segments and JIT encryption
based on the User Agent of the
client.
1080 @ 5Mbps
720 @ 3Mbps
576 @ 1.4Mbps
360 @ 1Mbps
240 @ 600kbps
AAC Stereo
Dolby Digital
MPEG-TS
5 video
2 audio
2-sec. segments
ENCODERSWINDOWS PC & MAC
ANDROID & IOS
HDS + Access
5 video
1 audio
6-sec. segments
HLS + PlayReady 2.2
4 video
1 audio
10-sec. segments
HLS + PlayReady 1.5
4 video
2 audio
2-sec. segments
1080 @ 5Mbps
Dolby Digital
Dolby Digital
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
Best practices and well architected - Video Solutions
Serverless content workflow
Get started with a pilot project – AWS Media specialists support
Show you how to build and iterate through your use case
138
GETTING STARTED
https://aws.amazon.com/answers/media-entertainment/
© 2017, Amazon Web Services, Inc. or its Affiliates. © 2017, Amazon Web Services, Inc. or its Affiliates.
• Part of AWS Elemental, Global Services is focused on delivering market leading
solutions on top of AWS & Elemental product offers.
• We are not a substitute for product, rather we deliver capabilities on top of core
product to extend, integrate, configure, deploy, and provide “System Integrator” like
solutions for our customers.
• We enable our customers to take our building blocks and deliver a solution to
fit their needs.
• Building Blocks are forms of AWS native services, Elemental suite of products and
3rd Parties.
GLOBAL SERVICES – WHO WE ARE
© 2017, Amazon Web Services, Inc. or its Affiliates.
THANK YOU
© 2017, Amazon Web Services, Inc. or its Affiliates.
MEDIA MICRO-SERVICES
© 2017, Amazon Web Services, Inc. or its Affiliates.
Amazon EFS
File
Amazon EBSAmazon EC2
Instance Store
Block
Amazon
S3
Amazon
Glacier
Object
Data Transfer (Ingest/Egress)
AWS Direct Connect AWS Snowball ISV Connectors Amazon Kinesis
Firehose
S3 VPC
EndPoint
AWS Storage
Gateway
S3 – Infrequent
Access
Events
S3 Event
Notifications
S3 Transfer
Acceleration
STORAGE
AWS STORAGE SERVICES MATURITY
ATTN:’s Journey to the
Cloud © 2017 Digital ReLab LLC
&
About Us (Digital ReLab and ATTN)
© 2017 Digital ReLab LLC
ATTN: Scaling (Choosing Starchive &
AWS)
© 2017 Digital ReLab LLC
Business Context:
1. Scaling content creation2. Desire to enable greater fidelity3. New types of content creation4. Desire to centralize and optimize
storage5. Approaching half a billion video
views monthly
ATTN:’s Situation
© 2017 Digital ReLab LLC
Scaling higher-quality content means:
1. More valuable B-roll to move to the cloud and monetize2. More A-Roll to publish3. Fast, reliable, and flexible transcoding is necessary4. Easy search and discovery is critical – requiring better metadata standards5. Future proofing to leverage “best-in-class” AI is critical6. “Increased fidelity = larger files” – necessitating scalable storage
Power Of Starchive Leveraging AWS
© 2017 Digital ReLab LLC
• Sitting on hard drives• Disparate data silos• Lost and duplicate files
• Making and Saving Money via:
• New Publishing channels
• Licensing• New Products• Greater Efficiency
• Transcoding• Metadata Probing• New metadata from AI• Search• Playback• Distribution
Digital Content Digital
AssetsValue Creation
Starchive in the Cloud (AWS)
© 2017 Digital ReLab LLC
Enabling New Tools in Media Transformation
© 2017 Digital ReLab LLC
The Media Transformation stage in the pipeline allows ATTN: to:
1. Transcode, probe metadata, connect business logic
2. Run AWS Rekognition3. Leverage new AI Tools via Marketplace4. Re-process files with updated AI
Results
With Starchive & AWS, ATTN: can now:
1. Move both A-Roll and B-Roll to the cloud more efficiently2. Automatically transcode files for easier publishing3. Make use of surfaced metadata for search and archiving4. Automate the adding of business logic data to all video as
it is created5. Pipeline new data transformation 6. Stay ahead of the curve by leveraging new AI as tools
emerge
© 2017 Digital ReLab LLC
Lessons Learned
Q & A
© 2017 Digital ReLab LLC
Thank You
© 2017 Digital ReLab LLC
Deep Media Analysis Using Amazon AI
Konstantin Wilms, Principal Solutions Architect, M&E, AWS
November 7, 2017
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Media & Entertainment Cloud
Symposium | Los Angeles
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
L ive2VOD Enr i chment
L ive Sent iment Ana lys i s
Deep V ideo Ana lys i s
Fas te r than Rea l t ime In fe rence
Agenda
ASSET INGEST AMAZON S3 AWS LAMBDA
CLIENT APP AMAZON ELASTICSEARCH
AMAZON REKOGNITION
Mobile app uploads
the image to S3
A Lambda function is triggered
and calls RekognitionRekognition retrieves the image from S3 and
returns labels for the property and amenities
Other users can search properties by
landmarks, category, etc.
MEDIA ASSET
MANAGEMENT
S3TA or 3rd Party Asset
Ingest
Event-Based Processing
Via S3 Notifications
Lambda Extracts Frames
& Conforms Image FramesReturn Object, Scene & Context
Data for Images
Inject Asset Tags via REST API
Into specific fieldsSearchable Tags stored as bag of
words collections
Authenticated Users can
Search Assets via Custom/Web App
Metadata Enrichment using Amazon Rekognition
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Since an early flush of optimism in the 1950’s, smaller subsets of artificial intelligence - first machine
learning, then deep learning, a subset of machine learning - have created even larger disruptions.
DEEP
LEARNING
MACHINE
LEARNING
ARTIFICIAL
INTELLIGENCEEarly Artificial Intelligence
stirs excitement
to flourish
Machine Learning begins
drive AI boom
Deep Learning breakthroughs
1950’s 1960’s 1970’s 1980’s 1990’s 2000’s 2010’s
© NVIDIA
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
The 10,000ft Intro to Deep Learning
Raw Data Low Level Features Mid Level Features High Level Features
Result
Application
Components
Task
Identify a Face
Training
10-100M images
Network
~ 10 layers
1B parameters
Learning
~ 30 Exaflops
~ 30 GPU days
© 2016 NVIDIA
Input
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Services & Partners
Amazon
Polly
Life-like
Speech
Amazon
Rekognition
Amazon
Lex
Deep Image
Analysis
Conversational
Engine
Deep Learning
AMI
Multiple
Frameworks
Amazon
ML
Predictive
Analytics
Managed DIY
Partner
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
The Amazon AI Stack
Services
Platforms
Frameworks
Infrastructure
MXNet TorchCTKKerasTheanoCaffeTensorFlow
AWS Deep Learning AMI
Amazon ML ECSSpark & EMR Kinesis Batch
Vision Speech Language
GPU / FPGA ServerlessCPU IoT Mobile
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Choosing the right Instance Type for AI
Instance
Name
GPU
Count
vCPU
CountMemory Network EBS
p3.xlarge 1 8 61 GiB ~10Gbps 1.5 Gbps
p3.8xlarge 84 32 244 GiB 10Gbps 7Gbps
p3.16xlarge 8 64 488 GiB 25Gbps 14Gbps
P2 & P3: Distributed Training & InferenceHyper-scale performance on NVIDIA V100s
G3: Multi-User ModelingNVIDIA M60 GPUs, 16,384 cores
F1: High Speed InferenceXilinx Ulstrascale Plus, 6,800 engines
X1: Specialized AI/ML/DL128 vCPUs, 3,904 GiB RAM
P3 Instances Provide up to 1 Petaflop of mixed precision performance, and
125 Teraflops of single precision floating point
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Deep Learning (& AI) for Media – Glass to Glass
Playout &
DistributionFiltering & Quality
Control
Visual Effects &
EditingApplication & Filesystem
Texture & Asset Search
AnalyticsSentiment Analysis
Other Amazon AI
Services
(Lex, Polly)
DAM & ArchiveAuto-categorization
Metadata Augmentation
Digital Supply ChainTag on Ingest
Live and VOD Feature
Extraction
Celebrity Detection
PublishingValue Add
API-based services
OTTFiltering &
Quality Control
AcquisitionPre-processing
& optimization
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Building A Deep Learning Pipeline in the Cloud
Training
Version Controlled
Datasets
Amazon
S3
Amazon
API Gateway
Amazon
EC2
Amazon
S3
AWS Lambda
Amazon
EFSAWS
BatchAmazon
ECS
Amazon
Glacier
AWS
Snowball
Data
Scientist
Amazon
Mechanical Turk
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Training a Deep Learning Pipeline
Youtube-8M CelebAPlaces CIFAR-10/100
hybrid cloud processing is well suited to large datasets
Object, Network & Gateway
Storage Services
Frameworks, 3rd Party Enablement, & Industry Initiatives
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Managed Serv i ce , Or D IY?
Deep learning-based image recognition serviceSearch, verify, and organize millions of images
Object and Scene
DetectionFacial
Analysis
Face
Comparison
Facial
Recognition
Celebrity
Recognition
Image
Moderation
Amazon Rekognit ion
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
• Object & Scene Detection
Metadata Enrichment for
MAM/DAM/CMS/UGC services
• Facial Analysis
User Sentiment Analysis
• Face Comparison
Associate content with ‘non-celbrity celebrities’
• Content Moderation
Automatically scan & approve content
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Media Analysis using Rekognition
Maple
Villa
Plant
Garden
Water
Swimming Pool
Tree
Potted Plant
Backyard
Flower
ChairCoffee Table
Living Room
Indoors
Object and scene detection makes it easy for you to add features that search,
filter, and curate large image libraries.
Object & Scene Detection
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Demographic Data
Facial Landmarks
Sentiment Expressed
Image Quality
Brightness: 23.6
Sharpness: 99.9
General Attributes
Facial & Sentiment Analysis
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Detect explicit and suggestive contentRecognize thousands of famous individuals
Celebrity Tagging Content Filtering
Celebrity Recognition & Image Moderation
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
3rd Party Software
• AWS AI AMI
• OpenCV
• ImageMagick
• FFMPEG
• Many Others
AWS Services
• Amazon S3
• AWS Lambda
• AWS SQS & SNS
• Amazon DynamoDB
• Amazon EC2 / SPOT
AWS Partners
• Asset Management
• Media Workflow
• Content Processing
• Image Optimization
• Feature Extraction
Extending Rekognition
{
"FaceMatches": [
{"Face": {"BoundingB
"Height":
0.2683333456516266,
"Left":
0.5099999904632568,
"Top":
0.1783333271741867,
"Width":
0.17888888716697693},
"{
"FaceMatches": [
{"Face": {"BoundingB
"Height":
0.2683333456516266,
"Left":
0.5099999904632568,
"Top":
0.1783333271741867,
"Width":
0.17888888716697693},
"
{
"FaceMatches": [
{"Face": {"BoundingB
"Height":
0.2683333456516266,
"Left":
0.5099999904632568,
"Top":
0.1783333271741867,
"Width":
0.17888888716697693},
"
{
"FaceMatches": [
{"Face": {"BoundingB
"Height":
0.2683333456516266,
"Left":
0.5099999904632568,
"Top":
0.1783333271741867,
"Width":
0.17888888716697693},
"
{
"FaceMatches": [
{"Face": {"BoundingB
"Height":
0.2683333456516266,
"Left":
0.5099999904632568,
"Top":
0.1783333271741867,
"Width":
0.17888888716697693},
"
Decoupling
Amazon
SQS
Amazon
SNS
Amazon
Kinesis
Amazon Elastic
Transcoder
Media Processing
Amazon API
GatewayAWS Batch
Amazon
EC2 Amazon
ECS
Compute
Glue
AWS
Lambda
Storage
Amazon
EFSAmazon
S3
Persistence
Amazon
DynamoDB
Amazon
ES
AWS Elemental
Serv ice Integrat ion
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Determinist ic Response T ime
~500ms Object & Scene Detect ion
~1.5s Search for 1mil Face Col lect ion
moderation level = safe
Building Rich Metadata Indexes using Rekognition
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Name: You?
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Back to our Agenda …
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Live2VOD Enr ichment
Transport
• API calls have a cost – conform & extract frames
• Scene change detection vs. frame extraction
• Streaming & Byterange Processing for S3
• Open Source & Commercial Ingest
Media
• Image enhancement, extraction & stabilization
• Unsharp mask, deconvolution – CPU impact
• Frame Extraction Offsets - PTS extraction
• e.g. FFMPEG w/deshake, vidstab, OpenCV
Live & On-Demand Conformance
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Live Sentiment Analysis
Amazon RedshiftAmazon
Quicksight
Live Subject Camera Embedded Application
Amazon S3
Analyze Faces
Shoppers enter and
browse in retail store
In-store cameras capture
live images of shoppers
A Lambda function is triggered
and calls Rekognition Rekognition analyzes the image and returns
facial attributes detected, which include
emotion and demographic detail
Return data is normalized and
staged in S3 en route to
Redshift
Analysis
Reports Periodic ingest of data into
RedshiftRegular analysis to identify trends in
demographic activity and in-store
sentiment over time
Live Sentiment Analysis
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Deep Video Analys is
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Distributed Deep Learning Systems Architecture
LO
AD
BA
LA
NC
ER
S
ORKER #1CPU WORKERS
#NDocker(algorithm#1)
Docker(algorithm#2)
..
Docker(algorithm#n)
AP
I
SE
RV
ER
S
AP
I
SE
RV
ER
S
GPU WORKERS
Docker(deep-algo#1)
Docker(deep-algo#2)
..
Docker(deep-algo#n)
m4
m4
m4
x1
• B2B & B2C have differing
workflows - orchestrator, vs. fan
out APIs with host pinning
• SOA, Microservices, Serverless
• Fractional billing is a must for
fleet diversification when
training
• Training - centralized storage
• Inference – ‘shared nothing’
• Hardware-targetable Containers
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Faster Than Realt ime Inference
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
High Speed FPGA Inference
Standard
Training
FPGAZebra
Int16
/int8
Caffe: Inference
GPUcuDNN
FP32
Caffe: Learning
TrainingData
NNDescription
TrainedNN
TrainedNN
IncomingData
Results
Zebra for
Inference
Application
Primitive
CL Execution Region
DDR4 DIMM
16GiB
DDR4 DIMM
16GiB
AWS
Shell
Multi-Port DDR
Controller
Multi-Port DDR
Controller
Input Data Loader Input Data UnLoaderInput Data
Packetizer
Results Data
DePacketizerResults Data LoaderResults Data UnLoader
DMA Controller
AXI Slave
• Deterministic, velocity-based processing
• Real-time sentiment analysis
• AWS F1 FPGA w/Xilinx Ultrascale
• 976GB RAM for large models
• ~1200fps (unoptimized) inference
• Caffe & MxNet DL Framework
• Train on G2/G3, run on F1
• Transparent GPU to FPGA targeting
using HDK/AFI & DL framework
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
FPGA Advantages for Deep Learning
Un-Optimized Performance Performance with Tuning
Machine Learning on FPGAs
Jason Cong, Chancellor’s Professor, UCLA
Director, Center for Domain-Specific Computing
http://cadlab.cs.ucla.edu/~cong
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Where does Med ia Secur i ty f i t in ?
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Machine Learning-based Data Security
Machine Learning Service to help
customers prevent data loss in AWS
Security Goals
• Categorize new or unknown threats
based on known and theorized
examples
• High coverage (true positive volume)
• High Accuracy (few false positives)
• Adaptive
Understand Your Data
Natural Language Processing (NLP)
Understand Data Access
Machine Learning
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Why is this important for Media?
Data outside of the scope of asset
archives & content can be the largest risk
Features
• Discovery, Classification & Protection
• Basic vs. Predictive Alerting
• Per-data object risk level assessment
• Compliance, Disruption, Ransomware,
Privilege escalation, Permissions,
Suspicious Access, PII data exposure
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Key Takeaways
• Managed services can be used to eliminate
‘undifferentiated heavy lifting’, allowing for niche AI focus
• Many traditional ‘in the cloud’ paradigms map to deep
learning
• AI-based technology provides unique opportunities to
enhance existing media delivery services
• Design using hybrid modeling, in-cloud training, & micro-
services
• Utilize compute diversification across GPU, CPU & FPGA,
combined with Object Storage (S3) & Fractional Billing
(SPOT)
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Thank You!
Enhanced Content Workflows Using Amazon AI
Konstantin Wilms, Principal Solutions Architect, M&E, AWS
November 7, 2017
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Media & Entertainment Cloud
Symposium | Los Angeles
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Confidence Labels
99.2%
Animal
Dog
Chihuahua
98.6%
Food
Dessert
Muffin
97.9% Collage
Dog or Muffin?
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Algorithm Viability
OCR
Are you
feeling
lucky?
Perceptual
Hash
Not a
chance
Deep Logo
AnalysisBingo
Word or Logo?
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Since an early flush of optimism in the 1950’s, smaller subsets of artificial intelligence - first machine
learning, then deep learning, a subset of machine learning - have created even larger disruptions.
DEEP
LEARNING
MACHINE
LEARNING
ARTIFICIAL
INTELLIGENCEEarly Artificial Intelligence
stirs excitement
to flourish
Machine Learning begins
drive AI boom
Deep Learning breakthroughs
1950’s 1960’s 1970’s 1980’s 1990’s 2000’s 2010’s
© NVIDIA
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
The 10,000ft Intro to Deep Learning
Raw Data Low Level Features Mid Level Features High Level Features
Result
Application
Components
Task
Identify a Face
Training
10-100M images
Network
~ 10 layers
1B parameters
Learning
~ 30 Exaflops
~ 30 GPU days
© 2016 NVIDIA
Input
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Services & Partners
Amazon
Polly
Life-like
Speech
Amazon
Rekognition
Amazon
Lex
Deep Image
Analysis
Conversational
Engine
Deep Learning
AMI
Multiple
Frameworks
Amazon
ML
Predictive
Analytics
Managed DIY
Partner
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
The Amazon AI Stack
Services
Platforms
Frameworks
Infrastructure
MXNet TorchCTKKerasTheanoCaffeTensorFlow
AWS Deep Learning AMI
Amazon ML ECSSpark & EMR Kinesis Batch
Vision Speech Language
GPU / FPGA ServerlessCPU IoT Mobile
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Frameworks & Infrastructure
Services
Platforms
Frameworks
Infrastructure
MXNet TorchCTKKerasTheanoCaffeTensorFlow
AWS Deep Learning AMI
Amazon ML ECSSpark & EMR Kinesis Batch
Vision Speech Language
GPU / FPGA ServerlessCPU IoT Mobile
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Choosing the right Instance Type for AI
Instance
Name
GPU
Count
vCPU
CountMemory Network EBS
p3.xlarge 1 8 61 GiB ~10Gbps 1.5 Gbps
p3.8xlarge 4 32 244 GiB 10Gbps 7Gbps
p3.16xlarge 8 64 488 GiB 25Gbps 14Gbps
P2 & P3: Distributed Training & InferenceHyper-scale performance on NVIDIA V100s
G3: Multi-User ModelingNVIDIA M60 GPUs, 16,384 cores
F1: High Speed InferenceXilinx Ulstrascale Plus, 6,800 engines
X1: Specialized AI/ML/DL128 vCPUs, 3,904 GiB RAM
P3 Instances Provide up to 1 Petaflop of mixed precision performance, and
125 Teraflops of single precision floating point
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Why is this Important?
Amazon EC2 Compute & EBS block storage supports second-level billing.
Combined with EC2 SPOT Fleet, this provides a up to ~90% cost savings over on-demand.
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Deep Learning Compute
AWS Deep Learning AMI
MXNet
Torch
CTKKeras
TheanoCaffe
TensorFlow
Amazon EC2
AnacondaIntel MKL
CUDA+cuDNN Python 2+3
Caffe2
• One Click launch
• Machine Image or Stack-based
• Single node or distributed
• GPU, CPU (& FPGA)
• NVIDIA & Intel acceleration
• Anaconda Data Science Platform
• Python w/ AI/ML/DL libraries
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Building A Deep Learning Pipeline in the Cloud
Training
Version Controlled
Datasets
Amazon
S3
Amazon
API Gateway
Amazon
EC2
Amazon
S3
AWS Lambda
Amazon
EFSAWS
BatchAmazon
ECS
Amazon
Glacier
AWS
Snowball
Data
Scientist
Amazon
Mechanical Turk
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Training a Deep Learning Pipeline
Youtube-8M CelebAPlaces CIFAR-10/100
hybrid cloud processing is well suited to large datasets
Object, Network & Gateway
Storage Services
Frameworks, 3rd Party Enablement, & Industry Initiatives
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Large Scale Document Analysis
• NLP Topic Modeling @ Clemson University
• 533,560 Documents, 32,551,540 Words
• 1.1 million vCPUs over ~3hrs
• EC2 Spot, Single AWS Region
• SLURM scheduler - overlay
virtual workflow automation
• Per second billing for
EBS & EC2
17 years of computer science journal abstracts and full text papers
from the NIPS (Neural Information Processing Systems) Conference (2,484 documents and 3,280,697 words)
Provisioning & Workflow Automation Framework
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Deep Learning (& AI) for Media – Glass to Glass
Playout &
DistributionFiltering & Quality
Control
Visual Effects &
EditingApplication & Filesystem
Texture & Asset Search
AnalyticsSentiment Analysis
Other Amazon AI
Services
(Lex, Polly)
DAM & ArchiveAuto-categorization
Metadata Augmentation
Digital Supply ChainTag on Ingest
Live and VOD Feature
Extraction
Celebrity Detection
PublishingValue Add
API-based services
OTTFiltering &
Quality Control
AcquisitionPre-processing
& optimization
Deep learning-based image recognition serviceSearch, verify, and organize millions of images
Object and Scene
DetectionFacial
Analysis
Face
Comparison
Facial
Recognition
Celebrity
Recognition
Image
Moderation
Amazon Rekognit ion
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Determinist ic Response T ime
~500ms Object & Scene Detect ion
~1.5s Search for 1mil Face Col lect ion
moderation level = safe
Building Rich Metadata Indexes using Rekognition
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Name: You?
ASSET INGEST AMAZON S3 AWS LAMBDA
CLIENT APP AMAZON ELASTICSEARCH
AMAZON REKOGNITION
Mobile app uploads
the image to S3
A Lambda function is triggered
and calls RekognitionRekognition retrieves the image from S3 and
returns labels for the property and amenities
Other users can search properties by
landmarks, category, etc.
MEDIA ASSET
MANAGEMENT
S3TA or 3rd Party Asset
Ingest
Event-Based Processing
Via S3 Notifications
Lambda Extracts Frames
& Conforms Image FramesReturn Object, Scene & Context
Data for Images
Inject Asset Tags via REST API
Into specific fieldsSearchable Tags stored as bag of
words collections
Authenticated Users can
Search Assets via Custom/Web App
Metadata Enrichment using Amazon Rekognition
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Service Enhancement using Amazon Lex + Polly
AMAZON LEX API GATEWAY + AWS LAMBDA
AMAZON VPC
API GATEWAY + AWS LAMBDA
User Requests
Transcode Job
Status
Lambda Recognizes
Request
Lex Responds with
Job Completion ETA3rd Party SaaS3rd Party Stovepipe
Connector
Internal Transcoding
Service
Transcode Cluster
Job Inspection
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Key Takeaways
• Running Deep Learning Infrastructure is hard
• Managed services can be used to eliminate ‘undifferentiated
heavy lifting’, allowing for niche AI focus
• AI for media is a cross-functional tech undertaking
• Many traditional ‘in the cloud’ paradigms map to deep learning
• AI-based technology provides unique opportunities to enhance
existing media delivery services
• Design using hybrid modeling, in-cloud training, & micro-services
• Utilize compute diversification across GPU, CPU & FPGA,
combined with Object Storage (S3) & Fractional Billing (SPOT)
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Marinus Analytics provides law
enforcement with tools, founded in
artificial intelligence, to turn big data
into actionable intelligence. The
Marinus flagship software, Traffic
Jam, is a suite of tools for use by
law enforcement agencies on sex
trafficking investigations.
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Thank You!
© 2016, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Dan Romuald Mbanga
Business Development Manager
AI Platforms and Engines
An Overview of AI in M&E
@dmbanga
Thank You
A Flywheel For Data
More Data Better Analytics
A Flywheel For Data
More Data Better Analytics
Better Products
A Flywheel For Data
More Data Better Analytics
Better ProductsMore Users
A Flywheel For Data
More Data Better Analytics
Better ProductsMore Users
Click stream
User activity
Generated content
Purchases
Clicks
Likes
Sensor data
A Flywheel For Data
More Data Better Analytics
Better ProductsMore Users
Click stream
User activity
Generated content
Purchases
Clicks
Likes
Sensor data
Object Storage
Databases
Data warehouse
Streaming analytics
BI
Hadoop
Spark/Presto
Elasticsearch
A Flywheel For Data
More Data Better Analytics
Better ProductsMore Users
Click stream
User activity
Generated content
Purchases
Clicks
Likes
Sensor data
Object Storage
Databases
Data warehouse
Streaming analytics
BI
Hadoop
Spark/Presto
Elasticsearch
Artificial
Intelligence
A Flywheel For Data
More Data Better Analytics
Better ProductsMore Users
Click stream
User activity
Generated content
Purchases
Clicks
Likes
Sensor data
Object Storage
Databases
Data warehouse
Streaming analytics
BI
Hadoop
Spark/Presto
Elasticsearch
Artificial
IntelligencePinpoint
AI: TL;DR
Fast Compute
Ubiquitous Data
Advanced Learning Algorithms
Artificial Intelligence At AmazonThousands Of Employees Across The Company Focused on AI
Discovery &
Search
Fulfilment &
Logistics
Enhance
Existing ProductsDefine New
Categories OfProducts
Bring Machine
Learning To All
• Applied Research
• Core Research
• Alexa
• Demand Forecasting
• Risk Analytics
• Search
• Recommendations
• AI Services | Rek, Lex, Polly
• Q&A Systems
• Supply Chain Optimization
• Advertising
• Machine Translation
• Video Content Analysis
• Robotics
• Lots of Computer Vision..
• NLP / NLU
Just a few Deep Learning Use Cases at Amazon…
~ 1997
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Recommendations & Ranking At Netflix
Personalized ranking,
page generation,
search, similarity, ratings
In 140 new countries,
simultaneously
Neural Style Transfer
Neural Style Transfer
Neural Style Transfer
Neural Style Transfer
http://mashable.com
Search
Watch
Listen
Play
Download
Purchase
Rate It
Review It
Sharing
Tagging
Bookmarking
Transactional Behavioral
Common M&E Analytic Use CasesConte
nt
•Top Content
•Engagement
•Play
Audience
•Acquisitions
•Churn
•Wher
Operations
•How much buffering
•Best CDN paths
Other
•Monetization
•Ad Spend
Latency: Realtime, NRT, Batch
Audience: Execs, Production, Content Programming, Marketing, Ops, App/Dev
Amazon AIIntelligent Services Powered By Deep Learning
More in
2017
Infrastructure CPU
Engines MXNet TensorFlow Caffe Theano Pytorch CNTK
ServicesAmazon Polly
Platforms
IoT
Speech
Mobile
Amazon
ML
Spark &
EMRKinesis Batch ECS
GPU
More in
2017
Chat
Amazon Lex
Amazon AI: Machine Learning In The Hands Of Every Developer
Amazon Rekognition
Vision
Automating Footage Tagging with Amazon
Rekognition
Built in 3 weeks
Index against 99,000 people
Saving ~9,000 hours a year in labor
More in
2017
Infrastructure CPU
Engines MXNet TensorFlow Caffe Theano Pytorch CNTK
ServicesAmazon Polly
Platforms
IoT
Speech
Mobile
Amazon
ML
Spark &
EMRKinesis Batch ECS
GPU
More in
2017
Chat
Amazon Lex
Amazon AI: Machine Learning In The Hands Of Every Developer
Amazon Rekognition
Vision
Solving Some Of The Hardest Problems In Computer Science
Learning Language Perception Problem
Solving
Reasoning
AWS Deep Learning AMI: One-Click GPU Deep Learning
Up To 40,000
CUDA Cores
Python 3 Notebooks
& Examples
(and others)(Volta at launch)
Scale for
Training
TensorFlow,
Apache MXNet
With Thanks To The Apache MXNet Open Source Contributors!
Apple
Bing Xu
MIT
Chiyuan Zhang
Qihoo 360
Yizhi Liu
Microsoft
Tianjun Xiao
Indiana
Tianjun Xiao
“Sounds great, Dan!
How can we get started?”
Direct Connect
Collect & Organize Data
KinesisStreams, Firehose,
Analytics
Upload
AWS Snowball Edge
100T Data Transport Device
With On-board Compute
Snowball Selfies: Customers Love AWS Snowball
Direct Connect
Kinesis
Snowball
Snowmobile
IoT
Database Migration
Streams, Firehose,
Analytics
Upload
Collect & Organize Data
Direct Connect
S3
Kinesis
Snowball
Snowmobile
IoT
Database Migration
DynamoDB
Elastic Search
Cloud Search
Glacier
RDS MySQL, PostgreSQL, MariaDB
Oracle, SQL Server
Aurora MySQL
PostgreSQL
Streams, Firehose,
Analytics
Upload
Collect & Organize Data
Direct Connect
S3
Kinesis
Snowball
Snowmobile
IoT
Database Migration
EC2
DynamoDB
EMR
Elastic Search
Cloud Search
Glacier
RDS MySQL, PostgreSQL, MariaDB
Oracle, SQL Server
Aurora MySQL
PostgreSQL
Streams, Firehose,
Analytics
Redshift & Spectrum
Athena
QuickSight
Upload
Amazon AI
Lambda
Collect & Organize Data
Direct Connect
S3
Kinesis
Snowball
Snowmobile
IoT
Database Migration
EC2
DynamoDB
EMR
Elastic Search
Cloud Search
Glacier
RDS MySQL, PostgreSQL, MariaDB
Oracle, SQL Server
Aurora MySQL
PostgreSQL
Streams, Firehose,
Analytics
Athena
QuickSight
Upload
Amazon AI
Lambda
Mech Turk
Glue
Collect & Organize Data
Redshift & Spectrum
Security & Encryption Integrated Across The Platform
Fine grained
access controls
Broad KMS
integrationServer-side
encryption
with CMK
Audit key
usage by
user & role
Import
keysPolicy
validation
& simulation
Use Case: Recommendations
The Concept
Capture Audience Signal Data
Create history of user and item preferences
Estimate similar users and items
Record these in Search Engine
Query Search Engine with User History
Enjoy recommendations!
Log
Storage
ETL
User
Interface
Serving
Layerusers
Recommend
Engine
users
Media platforms
Mobile
Apache MXNet
SearchPlayBuyRate
Recommendations
Step 1: Logs History Matrix
User1 Thing1
User2 Thing2
User3 Thing3
User2 Thing4
User5 Thing1
User1 Thing2
User1 Thing3
Mike
Jon
Mary
Phil
Kris
Logs History Matrix
Step 2: Estimate Similar Things
History Matrix
2 8
2 4
8
4
Item-Item Matrix
Step 3: Reduce to Interesting Pairs
2 8
2 4
8
4
Item-Item Matrix
LLR
Indicators
(“Items Similar To This….”)
Step 3: Reduce to Interesting Pairs
Indicators
(“Items Similar To This….”)
Items Similar To This
Step 4: Store Indicators in a Search Engine (BATCH)
Superman Highlander, Dune
Star Wars Raiders, Minority
Report
Highlander Superman
Mulan Home Alone, Mermaid
Star Trek …
… …
4587 223, 5234
748 5345, 235
12 8234
245 9543, 7673
3456 4587
… …
Index
Indicators
Step 5: Query Search Engine w/ User History
748 Star Wars
12 Highlander
245 Mulan
4587 Superman
3456 Star Trek
Query
“12”
5345
3456
12
Amazon AI Lab
Partner with ML engineers
from Amazon to:
Prototype ML with your data
Learn through hands-on
workshops & training
Build custom models
in production