---
title: "The Graviton Trade: Price-Performance vs. Migration Drag"
description: "AWS Graviton migration trades strong price-performance gains against real coordination costs. This map shows the benefit, the cost, and the conditions that decide when Arm adoption pays off."
---

# The Graviton Trade: Price-Performance vs. Migration Drag

**Type:** [Trade-offs](https://yomu.fyi/visual-explainers/category/trade-offs.md)

![A three-panel tradeoff diagram for AWS Graviton migration: benefits such as annual savings and lower CPU use, coordination costs such as arm64 build runners and multi-architecture images, and conditions such as opt-in scheduling and profiling first, connected left to right toward a decision cue.](https://yomufyi.s3.us-east-1.amazonaws.com/visual-explainers/4883ad64-e298-4e0b-8df9-f6e16d6dd7bc/image-95f1d9c9-1c92-4ac6-bc76-18dfb915fc12.webp)

> AWS Graviton migration trades strong price-performance gains against real coordination costs. This map shows the benefit, the cost, and the conditions that decide when Arm adoption pays off.

## Explanation

This explainer treats AWS Graviton adoption as a trade rather than a switch. The benefit side is measurable: a connected-vehicle platform saved roughly $600,000 a year, improved NGINX ingress price-performance by about 40 percent, and cut CPU consumption by 57 percent as more than 65 percent of the platform moved to Arm. The cost side is coordination, including native arm64 build runners, multi-architecture container images, and node provisioning changes that must exist before teams can move. The condition side is control: opt-in scheduling with taints and tolerations, workload-level cost governance, and profiling that finds the real bottleneck before the architecture is blamed.

## FAQ

### Why would a team move only part of a platform to Graviton?

Because scheduling can be opt-in. In the connected-vehicle example, Graviton nodes were marked with taints and tolerations so teams opted in and migrated at their own pace, and more than 65 percent of the platform had moved without a disruptive cutover.

### What must be in place before workloads can run on Arm?

Build and image pipelines usually come first: native arm64 CI runners and multi-architecture container images published as manifest lists let both architectures run from the same artifact. Node provisioning also has to handle the new instance types, for example through Karpenter.

### How do you know a Graviton migration is actually worth it?

Profile the workload before blaming the architecture. In one registry scaling study the real limits were storage saturation, a test client's GIL, and kubelet PID exhaustion rather than CPU, so the fix was capacity and test design. Pair that evidence with workload-level cost controls so spend stays attributable.

**Tags:** [ARM](https://yomu.fyi/topic/arm.md), [AWS Graviton](https://yomu.fyi/topic/aws-graviton.md), [CI/CD](https://yomu.fyi/topic/ci-cd.md), [Cost Optimization](https://yomu.fyi/topic/cost-optimization.md), [Graviton](https://yomu.fyi/topic/graviton.md), [Kubernetes](https://yomu.fyi/topic/kubernetes.md)

## Related reading

### [How BMW Group accelerated connected vehicle workloads on AWS Graviton](https://yomu.fyi/post/how-bmw-group-accelerated-connected-vehicle-workloads-on-aws-graviton.md)
- Company: [AWS](https://yomu.fyi/company/aws.md)
- Author: Tomislav Lozančić
- Published: Oct 29, 2026
- Type: Problem & solution

BMW Group’s Connected Vehicle platform, Orbit, powers digital services for more than 24.5 million vehicles worldwide. In 2024, the team began migrating from x86 to AWS Graviton Arm-based compute to advance economic and environmental goals. They adopted Karpenter for flexible node provisioning, extended GitHub Actions runners with Graviton for native arm64 builds, and used multi-architecture container images via manifest lists to support both architectures transparently. The migration achieved approximately $600,000 in annual infrastructure savings, a 40 percent price-performance improvement on the NGINX ingress controller, and a 57 percent reduction in CPU consumption while scaling efficiently during peak loads. Today, over 65 percent of the platform runs on Graviton, with an opt-in scheduling model using taints and tolerations that allows teams to migrate at their own pace without disruption.


### [Smarter cost controls for serverless](https://yomu.fyi/post/smarter-cost-controls-for-serverless.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Piyush Singh, Vijay Raja
- Published: Oct 29, 2026
- Type: Announcement

Databricks is extending cost control functionality to serverless compute for notebooks, jobs, and pipelines, matching the controls already available for serverless SQL and Model Serving. The new controls operate at the workload and user layer rather than the infrastructure layer, with three capabilities: attribute costs through tags on serverless compute objects, control access via entitlements on default interactive and automated compute objects, and limit spend with per-compute rate limits expressed as t-shirt sizes S, M, L, XL. Serverless Compute Access Controls are generally available, rate limits are in beta, and serverless compute tags are in beta. Monitoring via system tables and budget alerts is generally available, letting teams track consumption within hours. The full set of controls moves cost governance from machine-level caps to people- and workload-level policies, addressing runaway spend, wrong hands, and lack of accountability.


### [Sizing Red Hat Quay: Profiling components to find real limits](https://yomu.fyi/post/sizing-red-hat-quay-profiling-components-to-find-real-limits.md)
- Company: [Red Hat](https://yomu.fyi/company/red-hat.md)
- Author: Bala Tripura Bodapati
- Published: Oct 29, 2026
- Type: Tutorial

Red Hat Quay is a container image registry with multiple configurable components, and a large enterprise customer experienced pull failures under load across 160 nodes. The team developed a systematic profiling method to answer how many concurrent pulls Quay can handle, testing a specific configuration with 2 Quay replicas, NooBaa storage, and HAProxy. They identified three bottlenecks: NooBaa storage saturation, test client limitations due to Python’s GIL, and kubelet PID exhaustion. Scaling NooBaa from 4 to 6 endpoints more than doubled throughput and recovered from 49% failure to 100% success. Final recommendations for 2 vCPU Quay were 25 concurrent pulls with 0.22 img/s, scaling to 16 vCPU with 300 concurrency and 0.56 img/s, though the 16 vCPU ceiling reflects a single test client limitation.


### [Automating agentic AI Well-Architected reviews with the AWS Agentic AI Lens](https://yomu.fyi/post/automating-agentic-ai-well-architected-reviews-with-the-aws-agentic-ai.md)
- Company: [AWS](https://yomu.fyi/company/aws.md)
- Author: Dinesh Subramani
- Published: Oct 28, 2026
- Type: Problem & solution

Smarsh, a compliance platform, needed to review its first multi-agent workload against the AWS Well-Architected Framework’s Agentic AI Lens, which existed only as published documentation with 150 best practices. To avoid a multi-day manual effort, Smarsh built a custom Claude skill that treats the lens as structured data and automates evidence gathering from the workload’s GitHub repository. The skill uses a five-value status rubric, honesty rules, and a deterministic Python script to produce a formatted workbook with per-item status, evidence, gaps, and remediation. Smarsh validated the skill against two synthetic fixture repositories—a moderately mature workload and a deliberately rough proof of concept—achieving 16 of 16 verifiable expectations, compared to 8 of 16 without the skill; the evaluation suite ships inside the skill for ongoing regression testing. The resulting reviews enabled sharper AWS engagement by mapping specific best-practice gaps to focused working sessions with AWS Solutions Architects.


### [Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design](https://yomu.fyi/post/meta-andromeda-powering-the-next-generation-of-ads-retrieval-through-m.md)
- Company: [Meta](https://yomu.fyi/company/meta.md)
- Author: Haiyu Lu
- Published: Oct 28, 2026
- Type: Announcement

Meta Andromeda is Meta’s personalized ads retrieval engine, and its next generation uses model-system co-design to achieve near ranking-level performance at retrieval scale. The system scores over one billion ads per second, a 100x throughput increase over the 2024 system, and has delivered a 2.4% increase in conversions on Facebook since 2025. The key innovation is a retrieval mixture-of-experts (MoE) architecture that reduces inference-time FLOPs by ~300x compared to the 2024 model, enabling ranking-level matching quality. The entire retrieval pipeline is fused into a single GPU kernel, eliminating memory bandwidth bottlenecks and achieving over 10 billion ads per second on a single NVIDIA GH200 Grace Hopper Superchip. The co-design approach, developed in collaboration with NVIDIA, positions the architecture to absorb more model complexity without proportional infrastructure growth.
