Meta AI Creates a New Platform for Executing Cutting-Edge AI Models

MultiRay is Meta AI's new platform to run large-scale artificial intelligence models smoothly

A new Meta AI research has developed MultiRay, a new platform for executing cutting-edge AI models at a large scale to make AI systems more productive. With Meta AI's new platform, MultiRay, countless AI models can share the same input. 

Only a fraction of the processing time and resources are used for each model, minimizing the overall cost of these AI-based operations. By centralizing the business's computational resources in one model, AI accelerators can easily deploy and strategically trade between computing resources and data storage. The universal models in MultiRay have been fine-tuned to excel in a wide variety of applications. Machine learning models for various uses, like subject tagging of posts and hate speech detection, can be upgraded and refined by teams across Meta AI with the help of MultiRay. This method saves us time and effort more than having multiple teams construct huge end-to-end models independently.

MultiRay has fastened the accessibility to Meta's big core models by offloading calculations to particular hardware like graphics processing units (GPUs) and reducing the time and energy expended on recomputation by keeping frequently used data in memory (cache). Meta AI's new platform MultiRay presently drives over 125 use cases across Meta, supporting more than 20 million queries per second (QPS) and 800 billion daily queries.

MultiRay employs huge, foundational AI models to reflect the input in a more perfect way that supplies a point in a high-dimensional vector space. An embedding represents the input that is more amenable to machine learning. To simplify the processing of task-specific models, MultiRay provides an embedding of the input data (such as text and images) that can be consumed in place of the raw input. MultiRay's core models are trained to perform well on various tasks, including similarity and classification. Due to the need to convey additional information, our embeddings are large (several kilobytes in size).