Over the last while I have moved everything I run on AWS to Graviton, the ARM CPUs AWS designs in house: the EC2 instances behind my services and the ECS containers that make up most of the actual workload. The short version is that it was worth it, the bill went down, and almost nothing about how I write or run code had to change. The longer version has a few sharp edges worth writing down, plus some genuinely interesting silicon arriving now in Graviton5.

Why ARM in the Cloud

The pitch is not “ARM is faster,” it is price/performance. For the same dollar you generally get more useful work out of a Graviton instance than the equivalent Intel or AMD one, and AWS prices the instances lower on top of that. Set expectations correctly, though: a single Graviton core will not smoke a top-bin x86 core on a single-threaded benchmark. The win shows up across a fleet of web servers, API workers, queue consumers, and databases: more cores per dollar and steady throughput under real concurrent load. If you are chasing the lowest latency on one hot thread, measure first.

ARM vs Intel vs AMD in Practice

• Graviton is my default: best price/performance for general compute, web tiers, microservices, and most databases, at the lowest cost and power.
• AMD is the value x86 option when something genuinely needs x86 but not Intel specifically.
• Intel is a deliberate choice for AVX-512, Intel-specific acceleration, or vendors that only certify on Intel. It is the priciest of the three.

What actually decides it is rarely the chip and almost always the software. If everything you run is open source or built from your own source, ARM is a non issue. The moment a proprietary, x86 only binary shows up, that workload stays on x86 until the vendor ships an ARM build. So the real question per service is not “is ARM fast enough,” it is “does every binary in this container have an arm64 version.”

What Actually Bit Me

None of these were dealbreakers, but I wish I had batched them up front instead of finding them one service at a time.

The first one is the biggest mindset shift: your images must be multi-arch. Your base images, every layer, and your own build all need linux/arm64. I moved to docker buildx and publish multi-arch manifests so one tag resolves correctly wherever it runs:

docker buildx build \
    --platform linux/amd64,linux/arm64 \
    -t myrepo/myservice:latest --push .

Then ECS has to be told which architecture to use. The task definition’s runtimePlatform defaults to x86, so set it to ARM64 and run on Graviton capacity or Fargate:

"runtimePlatform": {
    "cpuArchitecture": "ARM64",
    "operatingSystemFamily": "LINUX"
}

When a service misbehaves after the switch, the first thing I check is whether the published tag actually contains both architectures. This one-liner saves a lot of guessing:

docker buildx imagetools inspect myrepo/myservice:latest
# look for linux/amd64 and linux/arm64 in the platform list

The rest are smaller traps worth knowing up front:

• Emulated CI is slow. Building arm64 on an x86 runner means QEMU, which is painful for anything compiled. You no longer have to: GitHub-hosted native arm64 runners are generally available, free for public repositories and now offered for private ones too, so the pipeline builds on real ARM hardware instead of emulating it.
• Native dependencies are the long tail. Interpreted code moves for free; compiled extensions (native npm modules, Python wheels, cgo, old vendored binaries) each need an arm64 build.
• Check your sidecars. Logging, metrics, and security agents need arm64 images too. The mainstream ones have them, but a sidecar that silently fails to start will ruin your afternoon.
• Managed services are the easy win. RDS, ElastiCache, and Lambda offer Graviton with no code change. Start there to build confidence.

Graviton5: What Changed

While I was mid migration, AWS moved the goalposts. The lineage shows the trajectory:

Gen	Cores	Microarch	ISA	Clock	Memory
Graviton2	64	Neoverse N1	Armv8.2-A	2.5 GHz	DDR4
Graviton3	64	Neoverse V1	Armv8.4-A	2.6 GHz	DDR5-4800
Graviton4	96	Neoverse V2	Armv9.0-A	2.8 GHz	DDR5-5600
Graviton5	192	Neoverse V3	Armv9.2-A	3.3 GHz	DDR5-8800

Graviton5, announced in December 2025 with M9g instances reaching general availability in mid 2026, is a bigger step than the version bump suggests. It doubles Graviton4 to 192 Neoverse V3 cores, built on a 3nm process as four chiplets of 48 cores each with up to ~420 GB/s between chiplets. It carries 192 MB of L3 cache (around five times the prior generation), 2 MB of L2 per core, DDR5-8800, and PCIe Gen6. It also ships a new Nitro Isolation Engine, which AWS describes as a formally verified hypervisor: the VM isolation property is mathematically proven, not just tested.

AWS quotes roughly 25% better compute than Graviton4 per core, with up to ~35% on web apps, up to ~35% on ML inference, and ~30% on databases from improved branch prediction. Take vendor percentages with the usual grain of salt and benchmark your own workload, but the architecture behind them is real. For densely packed container fleets the per core number matters less than doubling the cores while growing cache and memory bandwidth. The general purpose M9g and M9gd are out now; the compute optimized C9g and memory optimized R9g are the variants I am waiting on, since most of my fleet maps to those shapes.

Is It Worth It

For me, clearly yes. If your stack is containers and managed services built from source or mainstream images, the migration is mostly mechanical: make images multi-arch, set the architecture on your tasks, fix the native dependency stragglers. Start with managed services for an easy win, then move containers one service at a time so you can roll back cleanly. The one piece of advice I would give my past self: audit dependencies and sidecars for arm64 support before flipping anything. Ninety percent of the work is trivial. The other ten percent is one unmaintained binary you forgot about, and it is much nicer to find it on a spreadsheet than in a failing deployment.