Lightning Studio, An IOS For AI Development?

The iPhone was transformative, not only because it is a powerful computer in your pocket, but because it is a platform for a myriad of applications available at the touch of the finger. That’s the intuition behind Lightning Studio, a platform that brings together the plethora of new AI tools in a single interface.

Lightning AI, the company behind the widely-used PyTorch Lightning framework, has announced the availability of Lightning Studio, to streamline the development and deployment of AI products. The release also includes Thunder, a powerful new open-source source-to-source compiler for PyTorch designed to speed up training and serving of generative AI models across multiple GPUs.

Lightning Studio aims to address the challenges faced by developers in the AI industry by providing a unified interface for all AI development needs.

"Software 1.0 such as web apps, servers, etc., can be efficiently built on laptops. With software 2.0 – specifically AI – thousands of GPUs, terabytes of storage, and collaborative capabilities are needed,” Will Falcon, CEO of Lightning AI, explained. “Studio is a cloud-based virtual environment where AI researchers and developers can code on a browser from their laptops to develop and ship AI collaboratively."

The platform offers a range of features, including pre-built templates, seamless scaling from CPU to GPU, and the ability to leverage natively integrated tools or build custom ones. Developers can deploy their AI products anywhere, whether on their own cloud, Lightning AI's cloud, or on their local GPU cluster.

One of the key innovations in Lightning Studio is its app-based architecture. Falcon drew an analogy to the iPhone, saying, "Apple didn't build the flashlights. They didn't build the calculators, they didn't build Spotify, they didn't build Uber, they didn't build Venmo, they didn't build PayPal. All of these were third party products that were put together through the iPhone. So that's kind of what we do as well."

By integrating a variety of AI development tools, both third-party and in-house, into a single platform, Lightning Studio eliminates the need for developers to switch between multiple single-point solutions for tasks such as monitoring, training, serving, data preparation, and hosting apps.

Thunder, the new open-source compiler introduced as part of the release, is a culmination of two years of research on the next generation of deep learning compilers, built with support from NVIDIA. It is designed to efficiently train and serve the latest generative AI models across multiple GPUs, achieving up to a 40% speed-up compared to unoptimized code in real-world scenarios. These speed-ups can save weeks of training time and significantly lower training costs.

Thunder leverages best-in-class executors and software from NVIDIA such as cuDNN, nvFuser, and torch.compile, as well as OpenAI's Triton. It allows developers to use all executors simultaneously, enabling each executor to handle the mathematical operations it was designed for.

The Thunder team is led by Dr. Thomas Viehmann, best known for his early work on PyTorch, key contributions to TorchScript, and making PyTorch work on mobile devices for the first time. His expertise is expected to drive upcoming performance breakthroughs that will benefit the PyTorch and Lightning AI community.

"Our collaboration with Lightning AI, to integrate NVIDIA technologies into Thunder, will help the AI community improve training efficiency on NVIDIA GPUs and lead to larger and more capable AI models," said Christian Sarofeen, Director of Deep Learning Frameworks at NVIDIA.

Pricing for Lightning AI Studio is based on a tiered model, with a free level for individual developers, a pro level for engineers, researchers, and scientists, a teams level for startups and teams, and an enterprise level for larger organizations in need of enterprise-grade AI. Thunder itself is open-source under an Apache 2.0 license.

With the launch of Lightning Studio and Thunder, Lightning AI aims to revolutionize the way AI products are developed and deployed, making the process more efficient, collaborative, and accessible to a wider range of developers and organizations.

As the AI community embarks on training increasingly capable models spanning multiple modalities, these tools will enable developers to fully leverage their compute resources without having to hire their own systems software and optimization experts.