AWS boosts MoE RL training throughput by 40% with EFA and DeepEP on EKS
AWS outlines an architecture combining Amazon EKS, Elastic Fabric Adapter (EFA), and DeepEP to accelerate Mixture-of-Experts (MoE) reinforcement learning (RL) training. The setup addresses three key challenges: balancing rollout generation and policy training, managing heterogeneous compute and memory demands, and optimizing communication between nodes. MoE models, which use sparsity to scale…
Key points
- AWS’s architecture combines EKS, EFA, and DeepEP to cut MoE RL training bottlenecks by 40% on 48 P5en instances
- DeepEP replaces generic NCCL collectives with NVLink/EFA-optimized kernels for sparse, dynamic MoE token routing
- Spot Instances reduce rollout generation costs by tolerating interruptions without stalling tightly coupled policy training
AWS claims the new approach delivers 40% higher aggregate RL rollout throughput on 48 P5en instances (16 for training, 32 for inference) compared to a baseline Slime stack. The improvement comes from replacing generic NCCL all-to-all collectives with DeepEP’s topology-aware kernels, which route tokens via NVLink for intra-node and EFA for inter-node communication. The system also leverages Amazon EC2 Spot Instances for cost-efficient rollout generation, isolating policy training workers from interruptions. The blog post provides step-by-step setup instructions, including EKS cluster configuration, EFA plugin deployment, and job submission via TorchX. AWS emphasizes modularity: components like rollout generation, policy training, and storage can scale independently, reducing idle capacity and cost.
The story so far
3 episodes →- AWS boosts MoE RL training throughput by 40% with EFA and DeepEP on EKSthis story
Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput
AWS Machine Learning Blog · 25 September 2026
Loading the full article…
This text was published by AWS Machine Learning Blog and written by Ashvin Nihalani. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- OpenAI’s Astra and Anthropic’s Opus break two unsolved Enigma messages · 1 src
- OpenAI fires contractors for using AI to train its models · 2 src
- OpenAI’s research chief discusses AI safety and self-improvement goals · 1 src
- AWS shows how to train Qwen3-VL-8B with SkyRL on SageMaker HyperPod · 1 src
- Researchers unveil UpTCR model for TCR-antigen binding prediction · 2 src
Comments
via GitHub Discussions