AWS recently launched a new custom chip Graviton3, and an instance of Trn1 based on the Trainium chip. As the world’s largest cloud computing provider, AWS has revenue of $45.37 billion in 2020 and is also one of the largest buyers of data center chips. Since the acquisition of a start-up company called Annapurna Labs in 2015, AWS has been committed to developing its own custom chips for use on servers.
AWS Graviton3 processor is an AWS self-developed chip based on the Arm architecture. It is the third-generation product of the Graviton series since its launch in 2018. AWS said that compared to Graviton2, Graviton3 has increased its computational performance by 25%, and its floating-point performance and encryption performance are twice the original. The machine learning performance is three times that of the original, while energy consumption is reduced by 60%. Graviton3 also adds a new pointer authentication function to improve security.
Although AWS has no specific instructions, there are rumors that Graviton3 uses the Armv9 architecture and Neoverse N2 core, and supports DDR5 memory. Graviton3 will provide support for AWS EC2 C7g instances, suitable for computing-intensive workloads such as HPC, EDA, distributed analysis, and CPU-based machine learning inference. The EC2 C7g instance provides 30 Gbps of network bandwidth and Elastic Fabric Adapter (EFA) support and is currently in the preview stage.
Trainium is a self-developed cloud inference chip developed by AWS. It is used to train deep learning models. It will be released in 2020. This time it provides the latest Trn1 instance based on the Trainium chip. AWS said that it will enable applications such as image recognition, natural language processing, and fraud detection to achieve the best price-performance ratio.
In addition, the AWS EC2 M6a instance uses AMD EPYC 7003 series processors based on Zen 3 architecture. Compared with the previous generation of EC2 M5a instances, the cost performance is increased by 35%, and the cost is reduced by 10% compared to other EC2 instances based on x86 processors. It is understood that AWS plans to introduce more instances of AMD’s third-generation EPYC series processors in the future.