AWS Cloud Rendering: Global Customer Cases
Pradeep Pillai
AWS India
pillaip@amazon.com
Agenda
• India
• Animation/VFX - Rendering in India
– Major companies
– Major Work
• Rendering on AWS : Client Case Studies
– Client1
– Assemblage Entertainment Pvt. Ltd.
– SixSteps Pvt Ltd.
– Housing.com
• Q&A
Simulating 1 second of human brain activity takes
3,300,000 Compute Cores
World's Fastest Supercomputer
3,120,000 Cores
Cost :$390 Million
~ 39 years of computational chemistry
Estimate On-Prem Cost: 40Million
Novartis: HPC job on AWS
Results on AWS:
87000 cores
Time taken : 9 hours
Cost : $4200
India
Population 1.2 Billion
Area 3,287,240 km²
Capital New Delhi
Largest City Mumbai – Population: 13,922,125
Official languages Hindi and English, 22 others
GDP (nominal) $2.2 trillion
About India
Major Animation/VFX Studios in India
Some good work done in India
8.7/10 IMDb
Production Cost: >$40M
Collection till date: >$100M
Best VFX till date by India
Client Case Study 1:
Rendering on AWS
------- Studios (India): 10K Rendering on AWS
Profile of the client:
Largest CG/Animation/VFX studio
3DAnimation Visual Effects Studio
USA India
Animation Movies
• Special Venue Attraction
Some of their customers:
Before AWS:
• 5000+ servers
• Multiple Petabytes Tape and Storage boxes
• One of the largest buyer of EMC Isilon Storage and Intel Servers in India
• Rent out servers
- Rush job for a high value customer
- Limited by their internal DC capacity
- Latest Cores and Large Size server not available
- Tight budget and delivery timelines
Engagement and getting them to move to AWS..
CFO Engg
Domain expertisein Rendering
Proof of Concept
scale this
• Now standardized
• No more HW purchase
Meter: Sample Dashboard
• Manages complete cluster lifecycle
– Launches servers and Monitors utilization
– Provides burn rate and estimated cost
– Data-driven SPOT deployment
Clients Competitive Analysis
MS Azure >7 Times costlier than AWS
Current Architecture
300 Mbps link
AMI + AWS CloudFormation Template
Lustre Render Nodes
On-Premises Storage
Maya workstations
On-prem farm Storage Servers )
Architecture and workflow notes
• For a 10K render, on-prem farm ran 1K or 2K jobs to iterate
• 3D wireframes modeling done on-prem using Autodesk Maya
• Once iterations done, wireframes/inventorytransferred to AWS, and
production10K render job kicks off
• Used Asperato accelerate wireframecopy
• LeveragedLustre cluster and render farm scaled up in time
(contd…)
• A typical rendering wireframe/inventorywould be around 2 to 4 TB
• Rendered assets transferred from Lustre to on-premfor compositing
• Data in Lustre FS would grow up to about 32 TB
• Customer worked out an arrangement with Pixar for Renderman
licenses on cloud
AWS Solution
• Hybrid model
1200 c3.8xlarge& r3.8xlargeSPOT instances
high memory req lighting passes
Pixar’s Renderman Rhythm & Hues
• Internetpipe of 300 Mbps (Upload 3D wireframes& Download rendered frames)
• Theyuse ASPERA overit
Price Achieved:
< $0.01
per core/hour over a 7½ day period
• Inclusive of:
– Compute
– Storage
– Network
– VPN
– Data Transfer
3,578 16,496
163,364
79,983
64,147
411 303
Jan Feb Mar Apr May Jun July
Spend ($) on AWS
What Next for this client…
EFS
SPOT Fleet SPOT Block
Appstream
GPU enabled workstation Nice
• Direct Connect
Next large project with this client…
• New 24k Render
500 instances of 32-cores 140+ GB RAMs
1500-2000 instances at peak
SPOT Fleet
Multi-Region render
Client Case Study 2:
Assemblage
Entertainment
Current Architecture
100 Mbps VPN
AMI + AWS CloudFormation Template
Spot Render node
s
NFS server
On-Premises Storage
Maya –
Modeling
On-prem farm Storage Servers
Nuke –
Compositing
Architecture and workflow notes
• 3D wireframes modeling done on-prem using Autodesk Maya
• Wireframes copied to an NFS server on AWS
• Uses Modo for rendering &
• Muster as scheduler and for Renderfarm Mgmt
• Render farms spread across multiple regions
• Nuke for compositing
• At peak used 900+ instances c3.4xLarge
Client Case Study 3:
Rendering Fox India
Profile of the client:
online render farm
Rendering as a Service
• Web Based console
3ds Max, Maya, Softimage
Plugins
Use of AWS:
• Uses AWS Spot instances in the backend
• Web console uses AWS API to Auto-Provision
• Workloads: Various sizes, Configs, S/W on AWS
• Clients include
• Animation/VFX Studios,
• Design & Engg Firms
• Architecture Firms
• Auto Manufacturing
• Electrical & Electronics Industries
Client Case Study 4:
Housing.com
Profile of the client:
Online Real Estate portal
funded Startup
Game changing innovations
Rendering houses with furniture
• Virtual Walk thru
SPOT
Command Line, API, Cloud Formation Templates (CfnCluster)
Take Control using AWS
Questions & Answers
감사합니다
Thank You

Pillai Pradeep - Global Rendering Customer Cases :: AWS Rendering Seminar -

  • 1.
    AWS Cloud Rendering:Global Customer Cases Pradeep Pillai AWS India pillaip@amazon.com
  • 2.
    Agenda • India • Animation/VFX- Rendering in India – Major companies – Major Work • Rendering on AWS : Client Case Studies – Client1 – Assemblage Entertainment Pvt. Ltd. – SixSteps Pvt Ltd. – Housing.com • Q&A
  • 3.
    Simulating 1 secondof human brain activity takes 3,300,000 Compute Cores
  • 4.
  • 5.
    ~ 39 yearsof computational chemistry Estimate On-Prem Cost: 40Million Novartis: HPC job on AWS Results on AWS: 87000 cores Time taken : 9 hours Cost : $4200
  • 6.
    India Population 1.2 Billion Area3,287,240 km² Capital New Delhi Largest City Mumbai – Population: 13,922,125 Official languages Hindi and English, 22 others GDP (nominal) $2.2 trillion About India
  • 7.
  • 8.
    Some good workdone in India
  • 9.
    8.7/10 IMDb Production Cost:>$40M Collection till date: >$100M Best VFX till date by India
  • 10.
    Client Case Study1: Rendering on AWS
  • 11.
    ------- Studios (India):10K Rendering on AWS
  • 12.
    Profile of theclient: Largest CG/Animation/VFX studio 3DAnimation Visual Effects Studio USA India Animation Movies • Special Venue Attraction
  • 13.
    Some of theircustomers:
  • 14.
    Before AWS: • 5000+servers • Multiple Petabytes Tape and Storage boxes • One of the largest buyer of EMC Isilon Storage and Intel Servers in India • Rent out servers - Rush job for a high value customer - Limited by their internal DC capacity - Latest Cores and Large Size server not available - Tight budget and delivery timelines
  • 15.
    Engagement and gettingthem to move to AWS.. CFO Engg Domain expertisein Rendering Proof of Concept scale this • Now standardized • No more HW purchase
  • 16.
    Meter: Sample Dashboard •Manages complete cluster lifecycle – Launches servers and Monitors utilization – Provides burn rate and estimated cost – Data-driven SPOT deployment
  • 17.
    Clients Competitive Analysis MSAzure >7 Times costlier than AWS
  • 18.
    Current Architecture 300 Mbpslink AMI + AWS CloudFormation Template Lustre Render Nodes On-Premises Storage Maya workstations On-prem farm Storage Servers )
  • 19.
    Architecture and workflownotes • For a 10K render, on-prem farm ran 1K or 2K jobs to iterate • 3D wireframes modeling done on-prem using Autodesk Maya • Once iterations done, wireframes/inventorytransferred to AWS, and production10K render job kicks off • Used Asperato accelerate wireframecopy • LeveragedLustre cluster and render farm scaled up in time
  • 20.
    (contd…) • A typicalrendering wireframe/inventorywould be around 2 to 4 TB • Rendered assets transferred from Lustre to on-premfor compositing • Data in Lustre FS would grow up to about 32 TB • Customer worked out an arrangement with Pixar for Renderman licenses on cloud
  • 21.
    AWS Solution • Hybridmodel 1200 c3.8xlarge& r3.8xlargeSPOT instances high memory req lighting passes Pixar’s Renderman Rhythm & Hues • Internetpipe of 300 Mbps (Upload 3D wireframes& Download rendered frames) • Theyuse ASPERA overit
  • 22.
    Price Achieved: < $0.01 percore/hour over a 7½ day period • Inclusive of: – Compute – Storage – Network – VPN – Data Transfer
  • 23.
    3,578 16,496 163,364 79,983 64,147 411 303 JanFeb Mar Apr May Jun July Spend ($) on AWS
  • 24.
    What Next forthis client… EFS SPOT Fleet SPOT Block Appstream GPU enabled workstation Nice • Direct Connect
  • 25.
    Next large projectwith this client… • New 24k Render 500 instances of 32-cores 140+ GB RAMs 1500-2000 instances at peak SPOT Fleet Multi-Region render
  • 26.
    Client Case Study2: Assemblage Entertainment
  • 28.
    Current Architecture 100 MbpsVPN AMI + AWS CloudFormation Template Spot Render node s NFS server On-Premises Storage Maya – Modeling On-prem farm Storage Servers Nuke – Compositing
  • 29.
    Architecture and workflownotes • 3D wireframes modeling done on-prem using Autodesk Maya • Wireframes copied to an NFS server on AWS • Uses Modo for rendering & • Muster as scheduler and for Renderfarm Mgmt • Render farms spread across multiple regions • Nuke for compositing • At peak used 900+ instances c3.4xLarge
  • 30.
    Client Case Study3: Rendering Fox India
  • 31.
    Profile of theclient: online render farm Rendering as a Service • Web Based console 3ds Max, Maya, Softimage Plugins
  • 32.
    Use of AWS: •Uses AWS Spot instances in the backend • Web console uses AWS API to Auto-Provision • Workloads: Various sizes, Configs, S/W on AWS • Clients include • Animation/VFX Studios, • Design & Engg Firms • Architecture Firms • Auto Manufacturing • Electrical & Electronics Industries
  • 33.
    Client Case Study4: Housing.com
  • 34.
    Profile of theclient: Online Real Estate portal funded Startup Game changing innovations Rendering houses with furniture • Virtual Walk thru SPOT
  • 35.
    Command Line, API,Cloud Formation Templates (CfnCluster) Take Control using AWS
  • 36.
  • 37.