Amazon has unveiled new chips for training and running AI models at its annual re:Invent conference. The first chip, AWS Trainium2, offers 4x better performance and 2x better energy efficiency than its predecessor. It can scale up to 100,000 chips in the AWS EC2 UltraCluster product and can train a 300-billion parameter AI language model in weeks. The second chip, Graviton4, is designed for inferencing and provides better compute performance, more cores, and more memory bandwidth than its previous generation. It also features encrypted hardware interfaces for enhanced security. Both chips aim to reduce reliance on GPUs, which are currently facing a shortage. Amazon has not provided specific release dates for the chips but expects them to be available next year.
