Paperspace Gradient machine learning platform is best known from extensive GPU support. They have recently partnered with Graphcore to provide new generation processors.
Paperspace was acquired by Digital Ocean on summer 2023.
Summary of Paperspace Gradient
Paperspace is a heavily investor backed company coming from the famous Y Combinator program. The company has been founded on 2018 in USA.
Paperspace has simplified their product catalog during the past two years. At the moment they have Core for computation and Gradient for machine learning and analytics.
Gradient has three main components. Notebooks for exploration, Workflows for training and finally the model Deployment.
This terminology is visible everywhere in the website and documentation which helps to understand their offering.
Paperspace pricing is based on monthly fee per user plus computation costs.
Cost per user
The free plan has limited resources and only public projects.
The first paid Pro tier brings mid-range computation instances available for 1-2 users for 8-12 $ per month per user. Projects are obviously private. Compute price will be added on top of the subscription fee.
Growth plan support a group of 5 users max.
For Enterprise edition you need to contact sales. Hosting on major cloud providers requires this plan.
When using Paperspace data centers, you pay according their regular computation prices .
When using your own on-premise computation resources you do not pay for compute.
When utilizing computing resources from other cloud vendors like AWS, Azure or Google Cloud you pay around 25% premium.
Paperspace runtime can be hosted on Paperspace servers, AWS, Azure, Google or on-premises.
Having their own infrastructure makes the company slightly different from other AI platforms providers. You do not need to rely on the hardware of the big vendors. Paperspace servers are located in California (US), New York (US) and Amsterdam (Netherlands).
The platform runs on Kubernetes on the background.
Parallel computation and IPUs on Paperspace
Gradient does not have ability for parallel computation at the moment. Parallel machine learning model training is coming soon.
Paperspace utilizes Intelligence Processing Units from a company called Graphcore. The IPU processor provide parallelism a bit similar than GPU processors.
Apparently the machine learning developers need the Poplar software kit to take advantage of Graphcore IPUs. Presumably the SDK would be comparable to CUDA for Nvidia.
Programming languages in Paperspace
Python. R can be selected as a runtime.
It should be possible to use your own Docker image to run jobs in any language or framework.