June 2022 Summaries

Grid.ai has rebranded as Lightning AI and introduced a unified platform intended to reduce fragmentation across the artificial intelligence development lifecycle, alongside announcing a $40 million Series B round led by Coatue with participation from Index Ventures. Founded in 2019 by PyTorch Lightning creator William Falcon and Luis Capelo, the company grew from the popularity of its open-source deep learning framework, which has been downloaded more than 20 million times. Lightning AI aims to let researchers, data scientists, and machine learning engineers build end-to-end machine learning systems in days rather than weeks or months by abstracting infrastructure concerns such as cloud computing, Kubernetes, distributed file systems, and fault-tolerant training. The company argues that disconnected tools for functions including data versioning, feature stores, experiment management, and hardware acceleration have made AI development costly and slow, and positions its platform as an integrated “operating system” for AI. The new identity also acknowledges the Grid platform’s history, while the funding is intended to support further expansion and simplification of the platform.

Lightning AI is an open-source, modular framework for creating and operating Lightning Apps across AI research, model training, deployment, and production MLOps workflows, with Grid.ai serving as its underlying infrastructure backend for scalable cloud-based training. Designed to reduce infrastructure management and engineering boilerplate, it aims to let researchers, data scientists, and engineers build, run, share, and scale AI applications using their preferred tools, including TensorBoard, Weights & Biases, and Optuna. Its Lightning Gallery provides curated reusable components and applications for capabilities such as training, serving, monitoring, and notifications, enabling users to assemble pipelines such as anomaly-detection systems. Example applications include PyTorch Lightning training and demos, cloud-hosted Jupyter notebooks, distributed hyperparameter optimization, and collaborative training across internet-connected machines. Lightning AI also offers near-zero-code components for converting existing Python modeling code into apps, while planned enhancements include running multiple local apps, multi-tenancy deployments, authentication, hot reloading, and pause-and-resume functionality.

Apple’s ARM-based M1 and M2 chips offer strong performance and energy efficiency, and PyTorch has announced plans to support their GPUs for deep learning workloads that previously ran mainly on CPUs. Early testing noted that convolutional neural networks may require substantially more memory on M1 GPUs and CPUs than on CUDA GPUs, creating potential swapping concerns, although the reported MacBook Pro runs remained below 80% memory utilization. To compare hardware fairly, the training tests used a batch size of 32 because Nvidia RTX 2080 Ti and GTX 1080 Ti GPUs were limited by 11 GB of VRAM, while a later M1 Pro test using batch size 64 was about 20% faster. Sebastian Raschka provides additional benchmarks and discussion in a linked article on running PyTorch on the M1 GPU.

SyntheticGestalt, a London- and Tokyo-based AI startup supported by academic and government organizations in the UK and Japan, develops machine learning systems for drug discovery, including novel drug candidates, enzymes, molecular representations, and protein–molecule interaction simulations. Its work requires processing potentially millions of chemical and protein text strings into feature-rich vectors and running extensive simulations involving thousands of configurations for many molecules, making scalable computing essential. After struggling to run multiple jobs concurrently with its earlier pipeline, the company adopted Grid, which enabled it to launch experiments in parallel, manage infrastructure and AWS usage more efficiently, monitor costs, and use Spot instances with automatic recovery. SyntheticGestalt reports that Grid reduced a 15,000-confirmation simulation stage from an estimated 40 days to one day and shortened a 100-by-100 experiment set from roughly 400 hours to four hours, while also expanding hyperparameter exploration and reducing the need to build internal computing infrastructure.