AIExplainer
Machine Learning Advanced 2 min read

What is a model parallelism?

A technique to split a large AI model across multiple computing devices

Model parallelism is a method used to train large artificial intelligence models by dividing the model into smaller parts and distributing them across multiple computing devices, such as graphics processing units (GPUs) or central processing units (CPUs). This allows for faster training times and increased model complexity.

Model parallelism is like a team of workers building a large house. Instead of having one worker build the entire house, the team splits the tasks, with some workers building the foundation, others building the walls, and others installing the roof. Each worker focuses on a specific part of the house, and when they're all done, the house is complete.

Google's BERT model uses model parallelism to train its large language model across multiple GPUs, allowing it to achieve state-of-the-art results in natural language processing tasks.

Model parallelism is used in deep learning to train large models that would otherwise be too big to fit on a single device. It's commonly used in applications such as natural language processing, computer vision, and speech recognition.

A common misconception is that model parallelism is the same as data parallelism, which splits the data across multiple devices. While both techniques are used to speed up training times, they serve different purposes and are used in different contexts.

Model parallelism has been used in various forms since the early 2000s, but it gained popularity with the rise of deep learning and the need to train large models.

tensor parallelism pipeline parallelism

Three products for different needs — explore what’s relevant to you.