Amazon Web Services (AWS) unveiled its cloud‑only custom processor "Graviton5."
AWS on Dec. 4 (local time) introduced Graviton5 at its annual AWS re:Invent 2025 event in Las Vegas as "a fifth‑generation chip that delivers the best price‑performance for a broad range of Amazon EC2 cloud workloads."
Graviton5 delivers up to 25% better compute performance than the previous generation while maintaining energy efficiency, enabling applications to improve run speed, expense, and sustainability at the same time. The Graviton5‑based EC2 M9g instance packs up to 192 CPU cores into a single package, achieving the highest core density in the EC2 lineup. With a design that shortens data travel distance, it reduces inter‑core communication latency by up to 33%, and it increases L3 cache capacity to five times that of the previous generation to boost responsiveness. Network and storage (EBS) bandwidth also increased by up to 15% and 20%, respectively.
On the security side, it applies the sixth‑generation Nitro Card of the AWS Nitro System and introduces the "Nitro Isolation Engine," which mathematically separates workloads from operators. By adopting a "zero operator access" architecture that prevents direct human access to EC2 servers, AWS said it strengthened reliability for sensitive data environments such as government, healthcare, and finance.
AWS said major customers including Adobe, Airbnb, Atlassian, SAP, and Siemens have already adopted the Graviton family of chips and are seeing performance gains and expense savings in video processing, gaming, semiconductor design, and ERP. SAP said SAP HANA Cloud OLTP query performance improved by up to 60% on Graviton5‑based M9g instances.
AWS said, "Graviton‑based instances are spreading so quickly that they now account for more than half of the new EC2 CPU capacity being added," and added, "With Graviton5, customers will be able to scale their workloads while reducing infrastructure expense and carbon footprint at the same time."
The M9g instance for general‑purpose workloads is currently available in preview, and the C9g and R9g instances for compute‑ and memory‑intensive workloads are scheduled to debut in 2026.