Free GPUs? Startup Hopes Free Is Right Price for GPU Cloud Service 

GPUs are famously expensive – high end Nvidia Teslas can be priced well above $10,000. Now a New York startup, Paperspace, has announced a free cloud GPU service for machine/deep learning development on the company’s cloud computing and deep learning platform. Designed for students and professional learning how to build, train and deploy machine learning models, the service can be thought of as an ML/DL starter kit that helps developers expand their skills and try out new ideas without financial risk.

Utilizing Nvidia Quadro M4000 and P5000 GPU’s and called “Gradient Community Notebooks,” the service is based on Jupyter notebooks and enables developers working with widely used deep learning frameworks, such as PyTorch, TensorFlow, Keras and OpenCV, to launch and collaborate on their ML projects. Similar to GitHub, Gradient Community Notebooks can be shared and “forked” into a user's own account while providing pre-loaded templates with various libraries, dependencies and drivers, the company said.

"GPUs are essential to ML development, yet the services available today are complex and prohibitively expensive for many developers," said Dillon Erb, CEO & Co-founder, Paperspace. "This is precisely why we created Gradient Community: to make GPU and ML development resources widely accessible and easy to deploy. Our focus on empowering developers with cutting-edge technology, and a means to collaborate, supports our mission to help every developer become an AI developer."

Training runs on the service are limited to six hours per session, though the number of six-hour instances per user is unlimited, Daniel Kobran, Paperspace COO, told us, explaining that those with enterprise-scale training runs of days or weeks would subscribe to a higher, paid tier on the Paperspace platform.

Kobran said the free service is intended to draw attention to Paperspace – “free” being a tried-and-true marketing power word. It’s also a “long game” marketing strategy to get students of AI – be they in college, at startups or even enterprise data scientists graduating from classical machine learning to deep learning – to become accustomed to using the Paperspace platform and then to evangelize it among colleagues. He likened Gradient Community Notebooks to Dropbox’s offer some years ago of 2GB of free storage as a way to gain market traction.

“We feel this is a really important time in the (ML/DL) technology lifecycle,” Kobran said, “getting mind share, we’re trying to attract as many people as possible so they can train on our platform, think of our platform as a de facto IDE (integrated development environment) for machine learning and data science and deep learning. The thinking is if we just get in front of more people, as they move on to bigger companies they’ll bring us with them into that organization. So it’s a longer bet…, it’s a classic bottum-up marketing model.”

Paperspace, which was founded in 2015 and has raised $23 million in venture capital, also hopes the service enables growth in the number of trained data scientists everywhere.

“We thought about how to make GPUs more accessible, and someone jokingly proposed that we charge zero dollars,” Kobran said, “and that became something we entertained, a way to crack open the broader market and let anyone in the world have access to powerful GPUs… We feel we can go after a much broader audience, (GPU cost) is a huge barrier to getting started with machine learning, and so we think this is a really good way to make it easier for anyone to get started with the technology.”

Besides access to free GPUs and CPUs, Paperspace said Gradient Community Notebooks offer a Jupyter-based collaborative environment, access to popular libraries, prebuilt models, and a project showcase, and it allows users to create a public profile page to share their bio and the work they're interested in. Users can start from scratch with a new notebook or leverage pre-configured projects from the Gradient ML-Showcase, a curated list of machine learning examples.