DNA Insider
繁中
Back
Official Source· AWS Artificial Intelligence· AI· Published Sep 29, 2026

AWS Reports 40% More MoE Reinforcement-Learning Rollout Throughput with EKS, EFA, and DeepEP

AWS describes a Mixture-of-Experts reinforcement-learning architecture combining Amazon EKS, EFA, DeepEP, and Amazon S3, and reports a 40% increase in aggregate rollout throughput for large-scale RLHF and GRPO training.

AWS Reports 40% More MoE Reinforcement-Learning Rollout Throughput with EKS, EFA, and DeepEP

Full Report

AWS has published an article on scaling Mixture-of-Experts (MoE) reinforcement learning on Amazon EKS. The architecture combines Elastic Fabric Adapter (EFA), DeepEP, and Amazon S3 for large-scale RLHF and GRPO training workloads.

AWS reports that the architecture increased aggregate reinforcement-learning rollout throughput by 40%. For teams generating rollouts as part of training, throughput is a measure of how quickly that work can be completed. The result offers a performance reference for the workloads described by AWS, but does not establish that other environments will achieve the same gain.

Why This Matters

Rollout throughput affects how quickly reinforcement-learning training workflows can generate data. AWS’s reported 40% increase gives teams considering EKS for large-scale RLHF or GRPO a concrete performance reference for the described architecture.

Evidence

Start with the primary evidence, then review supporting sources and context.

1 items

Primary Evidence

1
  • OfficialAWS Artificial IntelligenceDocumentOriginal

    Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput

    Ashvin Nihalani

    “Learn how to scale Mixture-of-Experts (MoE) reinforcement learning on Amazon EKS using Elastic Fabric Adapter (EFA) and DeepEP. This post presents an architecture that combines Amazon EKS, EFA, and Amazon S3 and increased aggregate reinforcement learning rollout throughput by 40% for large-scale RLHF and GRPO training.”

Recommended signals

Recommended by brand, topic, and recent related coverage

No more recommended signals right now

Official statement

Open original source
Back to Live FeedExplore more signals