Amazon Web Services has unveiled the next iterations of its custom-designed chips at its annual re:Invent conference in Las Vegas, introducing the Graviton4 and Trainium2 processors. The new chips represent significant advances in delivering the best price-performance and energy efficiency to power the variety of workloads AWS customers rely on.
Graviton4, the fourth generation of AWS’s Arm-based server CPU, provides up to 30 percent better performance than the previous Graviton3 chips. AWS claims Graviton4 offers 50 percent more cores and 75 percent more memory bandwidth. This boost will help Graviton4 maintain its reputation as the optimal choice for cost-efficient general purpose computing on Amazon EC2.
New R8g EC2 instances utilizing Graviton4 are aimed at memory-intensive applications like databases, analytics and caching. With up to 3x more vCPUs and memory capacity versus the current R7g instances, R8g will allow customers to analyze larger datasets faster and at a lower total cost of ownership.
On the artificial intelligence front, AWS introduced Trainium2, designed specifically for training the massive models enabling advances in generative AI. Trainium2 is predicted to achieve up to 4x faster training throughput and 3x greater memory capacity than the first Trainium silicon.
AWS will deploy Trainium2 chips within new EC2 Trn2 instances containing 16 chips each. Even more impressively, AWS plans EC2 UltraClusters of 100,000 individual Trainium2 processors interconnected via high-speed EFA networking. This extreme scale training infrastructure is expected to deliver exascale levels of performance, enabling previously unattainable AI models to be trained within weeks instead of months.
Major AWS customers like AI safety startup Anthropic expressed excitement around Trainium2’s potential to accelerate their work developing powerful but beneficial foundation models. Anthropic Co-Founder Tom Brown stated the new chip will be “at least 4x faster than first generation Trainium chips for some of our key workloads.”
By continuously advancing the silicon underlying its cloud, AWS aims to give customers more options and capabilities than ever to deploy virtually any application or workload efficiently. The announcements of Graviton4 and Trainium2 demonstrate AWS’s leadership in custom chip design and commitment to cloud innovation. Only time will tell what new frontiers these upgrades help organizations reach.
