AWS PCS now provides a production-ready Deep Learning AMI
Skip custom AMI builds — AWS now maintains a production-ready GPU AMI with EFA, Slurm, and PCS Agent pre-validated for instant cluster deployment.
View original announcement →Visual Summary
graph TD
A{{PCS-ready DLAMI}}:::announced
B(AWS Parallel Computing Service):::compute
C([NVIDIA GPU Drivers & CUDA]):::feature
D([EFA & Lustre Drivers]):::feature
E([Slurm Auto-Selection]):::feature
F(Amazon EFS):::storage
G(SSM Parameter Store):::storage
H((ML/HPC Users)):::external
I([PCS Agent]):::feature
H ==>|"deploys cluster"| B
B ==>|"launches nodes with"| A
A -->|"includes"| C
A -->|"includes"| D
A -->|"activates"| E
A -->|"manages via"| I
I -->|"connects to"| B
A -->|"mounts"| F
G -.->|"provides AMI ID"| A
classDef announced fill:#ff9900,stroke:#ec7211,color:#fff,font-weight:bold
classDef compute fill:#e3f2fd,stroke:#1565c0,color:#1565c0
classDef storage fill:#e8f5e9,stroke:#2e7d32,color:#2e7d32
classDef feature fill:#fff3e0,stroke:#e65100,color:#e65100
classDef external fill:#f5f5f5,stroke:#616161,color:#616161
What's New
AWS has launched the PCS-ready DLAMI, an AWS-maintained Amazon Machine Image designed specifically for running AI/ML training and HPC workloads on AWS Parallel Computing Service (AWS PCS). Built on the Deep Learning Base GPU AMI (Ubuntu 24.04), it bundles all critical infrastructure components — GPU drivers, EFA networking, Slurm, and PCS management tooling — pre-installed and compatibility-tested. This eliminates the need for customers to build and validate custom AMIs before deploying production-grade GPU clusters on PCS.
How It Works
- Base layer inheritance: The AMI inherits Ubuntu 24.04, NVIDIA GPU drivers, CUDA toolkit, EFA (Elastic Fabric Adapter) drivers, and the Lustre client directly from the upstream Deep Learning Base GPU AMI, ensuring a well-tested OS and driver stack.
- PCS-specific additions: On top of the base DLAMI, AWS adds the PCS Agent (cluster management daemon), Slurm for AWS PCS (multiple supported versions pre-installed), and EFS utilities for mounting Amazon EFS file systems.
- Automatic Slurm version selection: Multiple Slurm versions are bundled in the AMI; the correct version activates automatically at instance launch based on the cluster's configuration, removing manual version management.
- AMI discovery via SSM Parameter Store: The latest AMI ID for each architecture can be retrieved programmatically using AWS Systems Manager Parameter Store paths (e.g.,
/aws/service/pcs/ami/dlami-base-ubuntu2404/x86_64/latest/ami-id), enabling automation and IaC workflows. - Console-integrated selection: When creating or editing a compute node group in the AWS PCS console, users can select "PCS-ready AMIs" from a filtered dropdown that matches the instance type's architecture.
- Layered customization model: The AMI does not include AI/ML frameworks (PyTorch, TensorFlow, JAX) or math libraries; users add their application stack via shared file systems or by building a custom AMI on top of the PCS-ready DLAMI.
- Regular AWS-managed updates: AWS publishes new AMI versions whenever the upstream DLAMI or PCS components are updated, delivering ongoing security patches and driver refreshes without customer intervention.
- AMI metadata transparency: Each AMI's Description field encodes the source DLAMI version, PCS Agent version, supported Slurm versions, and EFS utils version for easy auditability.
Why It's Important
- Eliminates AMI build toil: Teams previously had to manually assemble, test, and maintain custom AMIs combining GPU drivers, EFA, Lustre, and Slurm — a time-consuming and error-prone process that this AMI eliminates entirely.
- Accelerates cluster deployment: With all infrastructure dependencies pre-validated, teams can go from zero to a running PCS cluster in minutes rather than days of AMI preparation and compatibility testing.
- Reduces operational risk: AWS-managed compatibility testing between NVIDIA drivers, CUDA, EFA, and Slurm versions reduces the risk of subtle driver conflicts that can cause hard-to-diagnose training failures at scale.
- Ongoing security posture: Regular AWS-published updates mean security patches and driver CVEs are addressed without requiring customers to maintain their own AMI pipelines.
- Lowers barrier for HPC/AI convergence: Organizations running both scientific HPC and large-scale AI training can use a single, unified AMI foundation, simplifying governance and standardization.
- Supports multi-architecture workloads: Native support for both x86_64 (e.g., GPU instances) and arm64 (e.g., Graviton-based instances) broadens the range of workloads and cost optimization strategies available.
How It's Different
- Versus self-managed custom AMIs: Unlike DIY AMIs, the PCS-ready DLAMI is AWS-owned and AWS-tested, meaning compatibility between all components (drivers, EFA, Slurm, PCS Agent) is guaranteed and continuously maintained.
- Versus the base Deep Learning AMI alone: The base DLAMI lacks PCS Agent, Slurm for PCS, and EFS utilities — making it unsuitable for PCS cluster nodes without additional configuration; the PCS-ready DLAMI adds exactly these missing pieces.
- Versus AWS ParallelCluster AMIs: AWS ParallelCluster uses its own AMI ecosystem; the PCS-ready DLAMI is purpose-built for the AWS PCS managed service and its Slurm integration model, and is not interchangeable.
- Versus SageMaker HyperPod: HyperPod uses its own lifecycle scripts and container-based approach; PCS-ready DLAMI targets Slurm-native HPC users who want more direct control over cluster topology and job scheduling.
- Automatic Slurm version activation: Unlike static AMIs that pin a single Slurm version, the PCS-ready DLAMI bundles multiple versions and selects the correct one at boot, making it forward-compatible with future PCS Slurm upgrades without an AMI rebuild.
- No additional cost: The AMI itself carries no incremental licensing or usage fee beyond standard EC2 and PCS pricing, unlike some third-party HPC AMI offerings.
When to Prefer It
- New AWS PCS cluster deployments: Any team standing up a new PCS cluster for GPU-based AI/ML training should start with the PCS-ready DLAMI rather than building from scratch, as it provides an immediately production-ready baseline.
- Teams without dedicated HPC sysadmin resources: Organizations that lack the expertise or bandwidth to maintain custom AMIs with up-to-date GPU drivers, EFA, and Slurm patches benefit most from AWS-managed updates.
- Large-scale distributed training workloads: When running multi-node jobs (e.g., LLM pre-training with PyTorch FSDP or Megatron-LM) where EFA and NCCL performance are critical, the pre-validated EFA driver stack reduces risk of networking bottlenecks.
- Regulated or compliance-sensitive environments: Teams that need auditable, versioned infrastructure can use the AMI Description metadata and SSM Parameter Store to track exactly which driver and agent versions are deployed.
- IaC-driven cluster automation: Teams using Terraform, CloudFormation, or CDK to manage PCS clusters can dynamically resolve the latest AMI ID via SSM, ensuring pipelines always deploy the most current, patched AMI.
- Hybrid HPC + AI workloads: Research teams running both traditional HPC simulations and GPU-accelerated ML jobs on the same cluster benefit from the unified foundation that supports both use cases.
- Rapid prototyping and benchmarking: When evaluating GPU instance types or distributed training frameworks, the PCS-ready DLAMI removes environment setup time, letting teams focus on benchmarking application-layer performance.
Availability
- GA status: Generally available as of June 1, 2026 — this is a production-ready release, not a preview.
- Regional availability: Available in all AWS Regions where AWS PCS is currently offered; no additional opt-in required.
- Architecture support: Available for both x86_64 (standard GPU instances such as P and G families) and arm64 (Graviton-based instances).
- Pricing: No additional cost for the AMI itself; customers pay standard AWS PCS and EC2 instance pricing only.
- Limitations — no application frameworks included: AI/ML frameworks (PyTorch, TensorFlow, JAX), compilers, and math libraries are not bundled; users must supply these via shared file systems or custom AMI layers.
- Limitations — PCS-specific: The AMI is designed exclusively for AWS PCS compute node groups and is not intended as a general-purpose DLAMI for standalone EC2 or other services.
- Discovery mechanism: Latest AMI IDs are available via AWS SSM Parameter Store and through the AWS PCS console's compute node group configuration UI.
- Reference architecture: The
awsome-distributed-aiGitHub repository (awslabs/awsome-distributed-ai) provides CloudFormation templates and training examples built on this AMI.
