AWS Graviton5 doubles cores, boosts performance 25% with new chiplet design

AWS has launched Graviton5, its fifth-generation custom processor, now generally available in the M9g and M9gd EC2 instances for general-purpose and agentic AI workloads. Graviton5 introduces a radically new four-chiplet architecture, housing 192 Neoverse V3 cores — double the core count of Graviton4. Each chiplet contains 48 cores, a DRAM controller, and a PCIe controller, connected via custom die-to-die links offering up to 420 GB/s of bandwidth. The processor moves to a 3nm process, enabling greater circuit density and faster on-chip communication. It also supports DDR5-8800 memory (the fastest in the cloud, per AWS) and PCIe gen6 interconnects.

Performance gains are substantial: each core delivers 25% better performance than Graviton4, and improved branch prediction in the Neoverse V3 core boosts real-world database workloads by up to 30%. The L3 cache has been expanded to 192 MB — over five times the previous generation — reducing memory latency significantly. Security is enhanced with the Nitro Isolation Engine, a formally verified hypervisor that enforces VM isolation with mathematical precision. Graviton now powers over 350 instance types, spanning web apps, microservices, analytics, ML inference, video encoding, and more.