AWS TRN1 Chips: Faster Machine Learning Model Training

AWS Unveils Trn1: Its Newest Machine Learning Chip

Driven by the increasing demand for tailored silicon solutions, Amazon has been actively developing custom chips to optimize performance for specific customer workloads.

The company initially launched the Inferentia chip in 2019, focusing on accelerating inference learning processes.

Subsequently, in the following year, Amazon introduced the Trainium chip, engineered specifically for the training of machine learning models.

Introducing the Trn1 Instance

Today, Amazon Web Services (AWS) has expanded upon its prior innovations with the release of its latest machine learning chip, the Trn1.

The announcement was made by Adam Selipsky during his inaugural keynote address at AWS re:Invent in Las Vegas.

“We are pleased to unveil the new Trn1 instance, powered by Trainium,” Selipsky stated to the re:Invent attendees. “This instance is projected to provide the most cost-effective performance for training deep learning models in the cloud, and the quickest performance available on EC2.”

High Bandwidth and Scalability

“Trn1 represents the first EC2 instance offering bandwidth of up to 800 gigabytes per second.”

“This capability is particularly advantageous for large-scale, distributed training scenarios involving multiple nodes.”

Selipsky highlighted its suitability for applications such as image recognition, natural language processing, fraud detection, and forecasting.

Furthermore, the chips can be interconnected to create even more potent performance through the implementation of “Ultra clusters.”

Ultra Clusters for Complex Models

“By networking these chips, we create Ultra clusters comprising tens of thousands of training accelerators.

These accelerators are interconnected with petabyte-scale networking, forming a powerful machine learning supercomputer.

This infrastructure is designed for the rapid training of highly complex, deep learning models containing trillions of parameters,” Selipsky explained.