AWS announced that EKS Auto Mode and the open‑source Karpenter now support Elastic Fabric Adapter and EC2 placement groups, enabling higher performance for distributed training and inference workloads.
AWS has expanded the capabilities of Amazon EKS Auto Mode and the open‑source Karpenter autoscaler by adding native support for Elastic Fabric Adapter (EFA) and EC2 placement groups.
Why EFA and Placement Groups Matter
EFA provides low‑latency, high‑throughput networking that is essential for tightly coupled distributed workloads such as deep‑learning training and large‑scale inference. Placement groups, meanwhile, allow EC2 instances to be positioned close together within the same hardware topology, further reducing network latency.
Integration with EKS Auto Mode
EKS Auto Mode now automatically selects instance types that support EFA and places them in placement groups when you enable the new options. This removes the need for manual configuration and helps ensure that pods receive the optimal network performance for their workloads.
Karpenter Enhancements
Karpenter, the community‑driven Kubernetes cluster autoscaler, has been updated to recognize EFA‑compatible instance types and to launch them within placement groups. The changes are delivered as part of Karpenter’s v0.30 release and are available through the standard Helm chart.
Both integrations support the same set of instance families that include EFA, such as the C6i, G5, and P4d families, and they work across all AWS regions where EFA is offered.
- Simplified configuration through EKS Auto Mode","Automatic placement group usage with Karpenter","Support for all EFA‑compatible instance families"
Customers can now achieve higher throughput for distributed training jobs, faster model inference, and more efficient scaling of high‑performance computing workloads without additional networking setup.
AWS announcement of EKS support for EFA and placement groups