This document summarizes a presentation given at Amazon about GPU instances on Amazon EC2. The presentation covered:
- An overview of Amazon's GPU instance offerings over time, including new G2 instances.
- How GPUs enable parallel processing that can accelerate applications.
- Examples of how Autodesk has used AWS GPU instances to enable remote desktop applications over the internet, allowing for improved collaboration.
- How the life sciences company Schrodinger uses AWS GPU instances to perform drug discovery workloads like free energy perturbation calculations, allowing them to scale simulations and reduce costs compared to owning hardware.
2. Agenda
Overview of GPU Instances
Gyuri Ordody with Autodesk: Evolution of CAD on AWS
Teng Lin with Schrodinger: Drug Discovery on AWS
Questions from audience (if there’s time)
26. Strategies
• Create new cloud services on server clusters
– Write or rewrite from scratch
• Move desktop technology to headless server
technology – EC2 instances and Amazon S3 as
backend
– Recreate UI functionality in the browser
• Deploy existing desktop apps in the cloud
– Reuse engine and GUI
26
28. Collaboration – Using the AWS Cloud
•
•
•
•
•
Access it anywhere
Access using any device
Seamless collaboration
Editing from anywhere
Data close to application
28
32. Autodesk Online Application - Architecture
Internet
Controller
EC2 Instance
…
Client
Region 1
User
Controller
EC2 Instance
Region N
Controller
EC2 Instance
Region 2
32
33. Autodesk Online Application - Architecture
Internet
Controller
EC2 Instance
Client
App Servers
Application Settings
User Data
EC2 Instances
Default User Data
Session Data
Connection
S3
SimpleDB
User
Custom Scaling
App Server AMI
Region
33
34. Application Remoting – Instance sharing
Home / Office 1
Autodesk Desktop Apps
Internet
…
Home / Office n
EC2 Instance
Internet
34
49. Drug discovery and development stages
• It takes $800 Million to $1 Billion and 10 to 15
years to develop a blockbuster
http://www.innovation.org/drug_discovery/objects/pdf/RD_Brochure.pdf
50. Simple facts about drug discovery
• Each development candidate has a value of
$50-100M
• But the overhead of producing these in
pharmaceutical company is $35-70M
– Success rate is only 1 in 3
– Thousands of molecules synthesized
• Pharmaceutical Industry needs to overcome the
innovation deficit in drug discovery process
52. Ligand-protein binding
• Altering receptor protein conformation, and
consequently changing biological functions.
• Binding affinity is critical for drug discovery
Yibing Shan etal Journal of the American Chemical Society, vol. 133, no. 24, 2011, pp. 9181–9183.
53. Free Energy Perturbation (FEP)
• Schrödinger’s FEP product
– Can predict binding affinity very accurately
• Key features
–
–
–
–
–
–
Better sampling algorithm
High quality Force Field
Perturbation network
Automated workflow
GPU support
Cloud capable
54. GPU is significantly faster
• Each edge takes 3 or more days on 96 cores
– Slow and unreliable due to cross node communication
– Perturbation network makes it even worse
120.0
109.8
Speed (ns/day)
100.0
86.4
80.0
60.8
60.0
DHFR
APOA1
40.0
26.3
20.0
18.5
15.2
21.0
4.9
0.0
8 x Intel Xeon X5672
GeForce GTX780
Amazon Tesla M2050 Amazon Geforce Grid
K520
55. How can FEP help drug discovery?
• Traditional drug design
– Takes weeks or even months to synthesize a compound
– Costs $1,000 to $5,000 per compound
– Synthesize thousands of compounds per project
• In-silico design using FEP
– Takes 72 GPU hours (~6 hours per calculation with 12 GPUs)
– Costs about $75, and the price keeps going down
– “want to do 1000 calculations per day”
56. Why AWS?
• Scalability
– Performed virtual screening using 50,000-core on AWS
• Security
• Price per FEP job
– It takes us two months to get GTX-780 cluster up running
$75.60
$80.00
$60.00
$40.00
$20.00
$16.61
$32.76
$32.01
$28.28
$17.62
$12.67
$3.78
$0.00
GTX-780 (50% Tesla K20 (70% Spot Instance
3-yr HEAVY
On-demand
util)
util)
CG1
reserved (100% Instance CG1
util) CG1
Spot Instance
3-yr HEAVY
G2
reserved (100%
util) G2
On-demand
Instance G2
57. FEP on cloud
• Next version will be cloud oriented
• Data will be processed and visualized on cloud
Auto Scaling
Mobile Client
Cluster
DB InstanceScaling GPU Cluster
Auto
Internet
Gateway
Client
Amazon EC2
VPN Gateway
Web Servers
VPN Connection
VPN
Connection
Traditional Server
Corporate
Data Center
58. Retrospective Study
Binding Affinity Prediction (kcal/mol)
-4
-5
-6
-7
-8
-9
-10
-11
-12
y = 1.07x + 0.652; R² = 0.599
(Linear regression to all ~150 ligands
across multiple systems)
-13
-14
-14
-13
-12
-11
-10
-9
-8
-7
Binding Affinity (kcal/mol)
-6
-5
-4
59. Blind test with company X on AWS
• 9 out of top 10 are active compounds
– Probability of achieving result this good is <1%
– “make half as many compounds”
– “save years of time on the project”
• Company X signed a88
contract after the test
90
Count
80
70
60
50
40
30
20
10
0
66
59
48
19
6
Highest
3
1
Binding Affinity
Lowest
60. Blind test with company Y on AWS
• 8 out of top 10 are the most active compounds
– Probability of achieving result this good is <1%
• Company Y wants us to provide a turn key solution
20
19
18
16
Count
14
12
10
10
10
8
6
4
4
1
2
8
0
Highest
1
1
Binding Affinity
Lowest
61. Prospective FEP with Company Z on AWS
• 1/3 of molecules are active instead of 1/7
• Company Z uses FEP on many projects
25
23
19
Count
20
15
10
Non-FEP
10
9
FEP predict to be active
FEP predict to be inactive
5
4
4
2
0
0
Highest
0
4
1 1
1
Binding Affinity
2
0 0
0
1
Lowest
62. Summary
• Computer aid drug design plays a critical role in
drug discovery
• Combining with GPU computing, accurate
modeling tools like FEP will accelerate the drug
discovery process
• Cloud is a viable solution for high performance
computing, in terms of pricing and scalability
• Amazon is the leader for GPU computing at
cloud
63. Please give us your feedback on this
presentation
CPN210
As a thank you, we will select prize
winners daily for completed surveys!