AWS Graviton 2-based instances vs. x86-based instances
Cloud giants, including AWS and GCP, have been developing custom chips for their data centers. As part of this effort, AWS acquired Annapurna Labs back in 2015 and successfully delivered custom chips with the lab. These chips include AWS Inferentia and AWS Graviton 2. The Graviton 2, announced in 2019, powers the next generation of EC2 core instances, AWS’s core computing service. With its lower cost and claimed higher performance, the AWS Graviton 2 has caught the attention of cloud users. This article will compare Graviton 2 instances to x86_64 instances in terms of cost, performance, and other considerations.
The Graviton 2, designed by AWS utilizing ARM Neoverse N1 Cores, comes with 30 billion transistors (7nm). It features 64 cores (not supporting SMT) and 64KB L1 caches and 1MB L2 cache per vCPU. One of the most significant differences when comparing the Graviton 2 instances to x86 instances (Intel Xeon and AMD EPYC) is that the Graviton 2 comes with 64 physical cores, while competitors count SMT logical cores as vCPUs.
Not only do the physical chip specs of the Graviton 2 draw attention, but the pricing structure has also garnered great interest from the cloud community. Providing the same vCPUs and memory, the Graviton 2-based EC2 instances are cheaper by 11% to 25% compared to EPYC or Xeon-based EC2 instances. Table 1 provides a detailed price comparison of different types of instances. Graviton 2-based instances offer price reductions of 25% compared to Xeon-based instances and 12% compared to EPYC-based instances. With the same vCPUs, memory, network performance, and similar caches and memory channels, the cost savings provided by Graviton 2-based instances are impressive.
Now, let’s discuss performance. Graviton 2-based instances exhibit enhanced performance depending on the workload types. Numerous benchmark tests have been conducted by both AWS and third parties, and the results are mixed. Video encoding and machine learning are typical fields in which Graviton 2-based instances have demonstrated better performance in general. The performance comparison appears to depend on various variables, such as workload types and code structures. Therefore, running tests before deciding which instance to use will be a crucial step.
There are a few considerations before employing Graviton 2-based instances. First, not all OSs are supported for Graviton 2-based instances. For instance, Windows, Oracle enterprise licenses, and older versions of Linux cannot be run on these instances. While ARM-based chips, including Graviton 2, are likely to support more OSs in the future due to other cloud providers developing their own ARM-based custom chips, it is not the case yet. Second, not all services support Graviton 2-based instances. For example, a popular security solution offered by Trend Micro only supports parts of its features. Lastly, specific functions, including AVX512 and x86-tuned codes, should be avoided.
In short, Graviton 2-based instances provide cost savings and enhanced performance if applied to the right workload. However, it is important to note that better performance and lower cost are not guaranteed. It is essential to check OS compatibility and dependency with other services and conduct performance tests. Once better performance is identified, deployment should be executed with blue/green deployment.
