What we call the public cloud was developed primarily to manage and deploy web servers. The target audience for these products is Dev Ops. While this is a massive and exciting market, the world of Data Science and Deep Learning is very different — and possibly even bigger. Unfortunately, the tools available today are not designed for this new audience and the cloud needs to evolve. This talk would cover what the next 10 years of cloud computing will look like.
Cloud Computing Was Built for Web Developers—What Does v2 Look Like for Deep Learning with Daniel Kobran
1. 1 / 22
Paperspace
DO NOT DISTRIBUTE
Paperspace
www.paperspace.com
Serverless AI for the future of intelligence.
2. 2 / 22
Deep Learning platform built for developers.
Infrastructure automation and software layer to build
intelligent applications.
Introduction
3. 3 / 22
A new generation of AI
developers require a rethinking
of tooling and workflows.
4. 4 / 22
Developers spend 75% of their time
managing infrastructure.
Why this matters
5. 5 / 22
The cloud was built for a different use-
case (web servers) and a different
audience (DevOps).
So what’s the underlying problem?
6. 6 / 22
Traditional Web Services Deep Learning
The DevOps ecosystem is rich
storage, CDN, deploy, monitor, VPC,
load balance, IPsec, CI/CD, DNS ...
data, notebooks, train, visualize,
collaborate, version, hyperparameters ...
?
+ 100s more
?
?
?
?
?
??
7. 7 / 22
The key to solving this problem
is finding the right layer of
abstraction.
8. 8 / 22
• BI
• Prediction
• Optimization
• Recommender systems
• Any heuristic
• GPUs
• Datastore
• Algorithms (CNNs, RNNs, ...)
• Frameworks (Pytorch,
TensorFlow, etc)
Put Uber/Facebook-grade
AI platform in the hands of every developer
InfrastructureBusiness Objective
There is a huge disconnect between modern business
objectives and the DL tools that can fulfill them.
Closing the Gap
9. 9 / 22
Paperspace abstracts powerful
infrastructure behind a simple
software layer making cloud
ML as easy as modern web
services.
10. 10 / 22
A complete platform for modern deep learning
Ingest → Train →
Analyze → Deploy
+
Manage, collaborate, share
• Fully-managed GPU infrastructure
• Unified dev experience
• 1 click Jupyter Notebooks/Lab
• API & language integrations
• ACL/team controls
11. 11 / 22
GRAI° Model building AI orchestration fabric
• Job queuing / management
• Cloud agnostic
• Accelerator architecture native
(GPU, FPGA, ASIC, TPU, etc)
• Unified compute
• Extensible
• Built on best practices
(containers, kubernetes, and data
policies)
Bare-metal
Frameworks Tooling
TensorFlow
VPS Private cloud (VPC) Public Cloud
GRAI° AI orchestration fabric
PytorchKeras PythonJupyter
Data
Quilt
Job queuing Unified compute Container
deployment
Data
management
Encryption /
key managment
12. 12 / 22
> import paperspace as ps
# Run job on GPU cluster
> ps({‘Type’: ‘TPU’, ‘container’: ‘TensorFlow’ ... })
Connecting modern ML and the cloud
by converting infrastructure into
code.
Raw compute is not sufficient.
Gradient Toolstack
GRAI Framework
Cloud Orchestration & Automation
Cloud Infra
Network . Storage . Compute
Cloudscale with a single line of code
14. 14 / 22
Trends:
1. Chip renaissance
2. Evolution of ML/AI in practice
3. Consolidation around best practices
Remarks from the trenches
15. 15 / 22
Chip renaissance
• Graphcore
• Cerebras
• Nervana
• Wave
• Google TPU
...
The big question today is whether
accelerator architectures will follow
commodity CPU x86 or lead to a
golden era for high-end, use-specific
hardware.
16. 16 / 22
Consumable API → Refit the Model → Model as core IP
The evolution of ML/AI in practice
2016 2018
• Clarifai
• AWS Rekognition
• Google Cloud Vision
• MS cognitive services
• Paperspace
• Algorithmia
• FloydHub
• ClusterOne
17. 17 / 22
• Containerization
• Jupyter
• Job runner architecture
• Pipeline
• etc.
Consolidation
around best
practices
18. Thank you.
Paperspace
20 Jay St. Suite 312
Brooklyn, NY 11201
hello@paperspace.com
www.paperspace.com
(718) 619 4325