AWS has provided a first look at its next-generation Graviton5 processor, a custom server CPU developed by Annapurna Labs for deployment across the company's cloud computing platform and AI inference infrastructure. The new processor represents a significant step forward in AWS's ongoing effort to develop in-house silicon tailored specifically for hyperscale data center workloads. Graviton5 adopts a chiplet-based architecture built around four compute dies manufactured using TSMC's advanced 3 nm process technology. Combined, the package delivers 192 Arm v3 performance cores, making it one of the highest core-count Arm server processors publicly disclosed. Each core includes 1 MB of dedicated cache, helping reduce memory access latency while improving performance consistency across heavily parallelized workloads. A major focus of the design is memory bandwidth. The processor integrates a 12-channel DDR5 memory subsystem supporting speeds up to DDR5-8800. AWS states that the platform is capable of delivering more than 800 GB/s of aggregate memory bandwidth. Such bandwidth is increasingly important for modern cloud services, including large-scale databases, virtualization, analytics, and AI inference applications that depend on rapid data movement between memory and compute resources.
Connectivity is provided through a 96-lane PCI Express Gen 6 root complex. The move to PCIe Gen 6 significantly increases available I/O bandwidth for accelerators, GPUs, networking hardware, and storage devices. As AI infrastructure continues to grow in complexity, faster communication between processors and attached devices becomes increasingly valuable. Despite being composed of four separate chiplets, Graviton5 operates as a fully coherent processor. AWS implemented a die-to-die interconnect delivering 420 GB/s of bandwidth, allowing all chiplets to share data while maintaining cache coherency across the package. This approach enables software to view the processor as a unified platform rather than a collection of independent compute dies. Each chiplet contributes an equal share of system resources.
A single die contains 48 Arm v3 performance cores, a three-channel DDR5 memory controller, and a 24-lane PCIe Gen 6 root complex. This balanced configuration helps ensure that compute, memory, and I/O resources scale consistently across the entire processor package. AWS indicates that Graviton5 is targeting performance improvements of up to 25 percent over AWS G4 instances currently powered by Intel Xeon Scalable Cascade Lake and AMD EPYC Genoa processors. While detailed benchmark results have yet to be published, the disclosed specifications suggest meaningful gains in compute density, memory throughput, and platform bandwidth. The processor underscores AWS's continued investment in custom silicon as cloud providers increasingly seek to optimize hardware around their own workloads. With 192 Arm cores, DDR5-8800 support, PCIe Gen 6 connectivity, and a high-bandwidth coherent interconnect, Graviton5 appears designed to address the growing demands of large-scale cloud and AI deployments.
| Specification | AWS Graviton5 |
|---|---|
| Process Technology | TSMC 3nm |
| Architecture | Arm v3 |
| Total CPU Cores | 192 |
| Chiplets | 4 |
| Cores per Chiplet | 48 |
| Cache per Core | 1 MB |
| Memory Channels | 12x DDR5 |
| Maximum Memory Speed | DDR5-8800 |
| Memory Bandwidth | Over 800 GB/s |
| PCI Express | 96 lanes PCIe Gen 6 |
| Die-to-Die Bandwidth | 420 GB/s |
| Target Performance Gain | Up to 25% |
Source: Ian Cutress (Twitter)