Amazon Web Services has unveiled the Graviton5, its fifth-generation custom Arm-based server processor, designed to challenge high-end AMD EPYC and Intel Xeon chips in cloud data centers. The new CPU features up to 192 Neoverse V3 cores, built on a 3nm-class process, making it the densest Armv9.2 processor available and the most powerful in the Graviton series to date. It is now available in preview for Amazon EC2 M9g instances, with compute-optimized C9g and memory-focused R9g variants set to launch in 2026. AWS claims the M9g instances powered by Graviton5 deliver up to 30% faster performance for databases, and up to 35% faster for web applications and machine learning workloads compared to the previous M8g generation. These gains stem from significant architectural improvements, including a redesigned internal layout that reduces inter-core communication latency by up to 33%—a critical achievement given the doubling of core count from Graviton4. One of the most notable changes in Graviton5 is the shift from a system-level cache (SLC) to a large, distributed L3 cache of approximately 180 MB. Unlike SLC, which sits outside core clusters and serves as a shared buffer for CPUs, GPUs, and accelerators, L3 cache is tightly integrated with core clusters, offering lower latency and better coherence. This change is essential at 192 cores, where a monolithic SLC would create bottlenecks due to increased mesh traffic and longer communication paths. The move to a multi-slice L3 allows hot data to remain close to the cores, improving performance predictability and scalability—key factors in multi-tenant cloud environments. The cache expansion is not just a scaling exercise. With a fivefold increase from Graviton4’s 36 MB SLC, and a claim of 2.6 times more cache per core despite doubling the core count, the architecture strongly suggests a distributed L3 design. This supports AWS’s goal of maintaining low latency and consistent performance across workloads. Graviton5 also features an upgraded memory subsystem, likely retaining a 12-channel DDR5 interface but operating at higher speeds—potentially above DDR5-5600. While this may result in lower per-core memory bandwidth compared to Graviton4, the larger L3 cache helps offset this by reducing DRAM access. Network bandwidth has increased by 15% on average, with up to double the throughput in top-tier configurations, and storage bandwidth via Amazon EBS has improved by about 20%. Security has also been enhanced with the introduction of the Nitro Isolation Engine, a formally verified isolation layer built into the sixth-generation Nitro Cards. This component uses mathematical proofs to guarantee workload separation from other tenants and from AWS operators, enforcing a zero-operator-access model. Customers will be able to review the formal proofs and implementation, adding transparency and trust—important for enterprises migrating from on-premises infrastructure. While AWS has not confirmed whether Graviton5 uses Arm’s official compute subsystem (CSS), the design appears to be a custom Annapurna Labs solution based on Neoverse V3 cores. The processor’s performance uplift is described conservatively, but the combination of 192 cores, 3nm process, and architectural refinements positions Graviton5 as a serious contender in the data center CPU market. In summary, Graviton5 represents a major leap in AWS’s in-house silicon strategy, offering a high-core-count, high-cache, high-efficiency Arm-based CPU that competes directly with top-tier x86 solutions. With improved performance, better scalability, and stronger security, it is poised to play a central role in AWS’s future cloud infrastructure.