TechPowerUp released the first images of Amazon’s in‑house server silicon, the Graviton5 processor from Annapurna Labs. Built on TSMC’s 3 nm node, the package stitches together four identical chiplets, each delivering 48 Arm V3 performance cores for a total of 192 cores. The design targets both cloud workloads and AI inference, positioning Amazon to replace Xeon and EPYC CPUs in its data centers and promises tighter workload isolation.

Each chiplet houses a three‑channel DDR5 memory controller and a 24‑lane PCIe Gen 6 root complex, aggregating to a 12‑channel DDR5‑8800 interface capable of over 800 GB/s bandwidth. A 420 GB/s die‑to‑die interconnect ensures full cache coherency across the four tiles, a rare feature for disaggregated designs. This bandwidth rivals high‑end accelerators and gives Amazon a clear path to squeeze more AI inference per watt. It also simplifies scaling by letting the same silicon serve mixed‑workload clusters.

Amazon Web Services claims the AWS Graviton5 will deliver roughly a 25 % uplift over its current G4 instances, which rely on Intel Cascade Lake and AMD Genoa Xeons. If the silicon lives up to the spec sheet, cloud customers could see dramatically lower cost per compute unit and tighter integration with Amazon’s own AI services, reshaping competitive dynamics in the hyperscale market.