Graviton2 versus Graviton4

Just a short post, because I thought this was pretty remarkable. Below, I have screenshots showing the CPU utilization of two AWS instances in us-west-2 which are running an identical workload.

They are running the CloudNativePG playground, which is a production-like learning and testing environment (all running virtually inside the single ec2 instance, which can be easily started and stopped or terminated and recreated). The standard CNPG playground setup consists of two Kubernetes clusters named kind-k8s-eu and kind-k8s-us. Each Kubernetes cluster contains a CloudNativePG cluster with HA between three postgres replicas running across three nodes/servers locally, and then there is cross-cluster replication from the EU cluster (primary) to the US cluster (standby).

What jumped out at me was the huge difference in CPU utilization! The Graviton2 instance runs maybe 40% utilization while the Graviton4 instance runs around 10% utilization.

I just now checked the AWS on-demand pricing page, and m6g.xlarge is $0.154/hr while m8g.xlarge is $0.17952/hour. That is a 16.6% increase in price, and for this particular workload it could be as much as a 300% increase in performance. At a fleet level, this should translate into significant cost saving opportunities if anyone adopted Graviton2 and if they are able to scale down overall instance counts based on better performance of newer generation chips.

Honestly, 40% utilization is technically fine for most of my own Kubernetes and Postgres experiments… but the 16.6% price increase is just low enough that I’ll probably start using the m8g instances anyway. 🙂

Things like this also underscore why it’s so hard to compare processors… how can we compare across different families, when we see differences like this between generations WITHIN a family?! Besides that, the total number of different processor choices we have today is overwhelming, taking into account all the different providers in the market. It’s a tough job.