AWS Unveils Custom AI Chips and Rapid ML Innovations at re:Invent
AWS hosts its first dedicated Machine Learning keynote at re:Invent 2020, highlighting over 250 new features and introducing custom training chips like AWS Trainium. The company also announces faster distributed training capabilities for Amazon SageMaker.
AWS hosts its first-ever dedicated Machine Learning keynote at re:Invent 2020, where Vice President Dr. Swami Sivasubramanian reveals an incredible pace of innovation with over 250 new ML and AI features released throughout the year. The presentation highlights AWS's dominant position in the cloud AI space, noting that 92% of cloud-based TensorFlow and 90% of cloud-based PyTorch workloads run on AWS infrastructure. This rapid release cycle averages less than two days per new feature, solidifying AWS as the most popular cloud platform among enterprise data scientists.
To accelerate future workloads, AWS announces the upcoming availability of Habana Gaudi-based Amazon EC2 instances in the first half of 2021, promising up to 40% better price performance than current GPU-based instances. Additionally, the company introduces AWS Trainium, a custom machine learning training chip designed in-house to deliver the most cost-effective ML training in the cloud. This custom silicon positions AWS to compete directly with other specialized AI accelerators like Google's Cloud TPU.
Amazon SageMaker also receives significant updates, with distributed training becoming generally available and enabling models to train up to 40% faster at no additional cost. This speed boost utilizes new data parallelism and model parallelism techniques to streamline complex deep learning workloads. Together, these hardware and software advancements demonstrate AWS's commitment to making machine learning faster, more accessible, and more cost-efficient for enterprise customers.