AWS Graviton

AWS Graviton. It is a family of server microprocessors designed by Amazon Web Services for its cloud computing platform. These processors are based on the ARM architecture, marking a significant shift from the industry-standard x86 designs provided by Intel and Advanced Micro Devices. The initiative aims to deliver cost-efficient and high-performance compute options for a wide range of Amazon Elastic Compute Cloud workloads.

Overview

The development of these processors stems from Amazon's 2015 acquisition of Annapurna Labs, an Israeli chip design company. This strategic move allowed Amazon Web Services to vertically integrate its hardware and software stack, following trends set by other hyperscale companies like Google with its Tensor Processing Unit and Microsoft with its Azure Maia initiatives. The first-generation processor was unveiled at AWS re:Invent in 2018, signaling a major commitment to custom silicon. This effort is central to the broader Amazon EC2 strategy of providing diverse instance types optimized for specific applications, from web servers to high-performance computing.

Architecture and Design

These processors are built on ARM's Neoverse platform, specifically designed for data center and infrastructure applications. The cores implement the ARMv8 and later ARMv9 instruction set architectures, offering features like Scalable Vector Extension support for enhanced machine learning inference. The design emphasizes energy efficiency and high core density, leveraging a custom system on a chip integration that includes DDR4 and DDR5 memory controllers and PCI Express connectivity. This integration reduces latency and power consumption compared to traditional discrete component setups, a principle also seen in Apple silicon designs for MacBook devices.

Processor Generations

The initial 2018 offering was followed by the significantly enhanced Graviton2, announced at AWS re:Invent 2019. Graviton2 featured 64 Neoverse N1 cores manufactured on a 7 nm process by TSMC, delivering a substantial performance leap. The third generation, Graviton3, was introduced in 2021, built on a 5 nm process and offering up to 25% better compute performance. It also introduced DDR5 memory support and new bfloat16 instructions. Subsequent specialized variants include the Graviton3E, optimized for high-performance computing workloads, and the Graviton4, announced at AWS re:Invent 2023 with 96 cores and DDR5 memory.

Performance and Benchmarks

Independent analyses and publications from Phoronix and SPEC have consistently shown these processors offer competitive or superior performance-per-dollar compared to equivalent x86-based Amazon EC2 instances. Benchmarks across common Linux workloads like NGINX, Memcached, and Redis demonstrate significant advantages for scale-out applications. In high-performance computing tests, such as those run with the OpenFOAM computational fluid dynamics software, Graviton3 instances have shown strong results. The performance gains are particularly pronounced in containerized environments managed by Kubernetes and for Java applications running on OpenJDK.

Use Cases and Adoption

These instances are widely adopted for running Linux-based applications, including web servers, application servers, microservices, and in-memory caches like Redis. Major software companies like Datadog, Elasticsearch, and MongoDB have optimized their distributions for the platform. Within the Amazon Web Services ecosystem, services like Amazon RDS, Amazon Aurora, Amazon ElastiCache, and Amazon EMR offer managed service options powered by these processors. The U.S. Air Force and Formula 1 are among notable enterprise adopters using them for simulation and data analytics workloads.

Comparison with x86 Instances

When compared to Amazon EC2 instances powered by Intel Xeon or AMD EPYC processors, the primary differentiator is the ARM architecture's inherent power efficiency, which translates to lower cost for comparable performance. This has created competitive pressure within the cloud computing market, influencing pricing strategies across Microsoft Azure and Google Cloud Platform. However, x86 instances retain advantages for certain legacy applications requiring specific instruction set extensions or with dependencies on proprietary software not yet ported to ARM. The choice often depends on workload-specific benchmarking, as seen in evaluations by the GitHub engineering team.