AWS Unveils New Custom Chips and SageMaker Tools for Cloud Machine Learning
AWS re:Invent 2020 highlights major machine learning updates, including custom training chips and new SageMaker features. These releases aim to reduce costs and simplify data preparation for production-scale deep learning.
AWS hosts its annual re:Invent 2020 conference virtually and dedicates an entire keynote to machine learning for the first time. This shift reflects the massive scale of cloud-based machine learning, as over 90 percent of TensorFlow and PyTorch workloads run on AWS infrastructure. The tech giant introduces several new tools designed to help developers train and deploy models more efficiently in the cloud.
The conference highlights significant hardware advancements, including AWS Trainium, a custom machine learning chip built specifically for cloud-based model training. AWS also unveils new EC2 instances powered by Habana Gaudi accelerators, which deliver up to 40 percent better price-performance than current GPU-based instances. Both hardware options integrate with the AWS Neuron SDK and Amazon SageMaker to provide cost-effective scaling for deep learning workloads.
On the software side, Amazon SageMaker receives a powerful new feature called SageMaker Data Wrangler. This tool allows users to process, transform, and visualize machine learning data with just a few clicks. Data Wrangler addresses the fact that data preparation consumes most of the time in machine learning projects by offering over 300 built-in transformations to simplify and accelerate this undifferentiated heavy lifting.