New Graviton5 memory instances: benchmark the working set before migrating

AWS has announced general availability of the Graviton5-based R9g and R9gd memory-optimized EC2 families. R9g uses EBS-backed storage, while R9gd adds local NVMe storage. The distinction matters for workloads that can benefit from low-latency local scratch space or caches.

The useful migration question is whether the application’s working set and bottlenecks match the new hardware. Databases and caches may benefit from memory, cache and bandwidth improvements, but a published family-level performance claim is not a guarantee for a specific query mix. Measure the same workload, dataset and latency target on both instance choices.

Check the complete Arm64 dependency chain before a move from x86: base images, native extensions, monitoring agents and vendor packages all count. A portable application language does not make every bundled binary portable. For existing Graviton workloads, still validate operating-system and agent compatibility.

Treat local NVMe capacity as instance-local storage with its documented lifecycle. Keep durable records in the appropriate persistent system, and test how the service rebuilds caches or scratch data after replacement. Compare total cost at the required throughput rather than comparing hourly prices alone.