From FP32 to INT8: The Science of Shrinking AI Models

Author(s): Harsh Maheshwari

Originally published on Towards AI.

Understanding quantization of neural network along with their implementation

This member-only story is on us. Upgrade to access all of Medium.

The training compute requirement for the famous AI models have become 45x in the last 10 years! The graph below contains data of this training compute requirement of notable AI models, over the years. Fitting a line on this data shows us that the requirement has increased 4.5 times per year.

Image from https://epoch.ai/data/notable-ai-models with CC license

In the context of AI models, training compute refers to the total computational power needed to train a model, which is proportional to the memory required. This includes both the storage for the model’s trainable parameters and the memory needed for the intermediate values generated during inference, which result from the input interacting with the parameters. As models grow larger, both the computational and memory requirements increase drastically.

For a computer, memory is ultimately measured in bits. One way to optimize memory usage is by changing how numbers are represented within the model. This technique, known as quantization, reduces the precision of numbers to save space and improve efficiency. Before diving into quantization, let’s first explore the different ways numbers can be represented in a computer.

The parameter values in a model are very commonly represented… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI