AWS released an updated infrastructure solution for scaling Mixture-of-Experts (MoE) reinforcement learning on Amazon Elastic Kubernetes Service (Amazon EKS), combining Elastic Fabric Adapter (EFA) and DeepEP to improve throughput by 40%. It is the company's first major update to its machine learning infrastructure since the introduction of EKS in 2017.
AWS reported a 40% increase in throughput for MoE reinforcement learning workloads, measured on large-scale asynchronous training tasks. That compares with earlier benchmarks that saw throughput constrained by communication overhead rather than compute.
The solution is built on Amazon EKS and targets large-scale reinforcement learning workflows, including RLHF and GRPO. Availability begins with general availability, initially for enterprise customers and research institutions.
"Scaling MoE reinforcement learning requires balancing heterogeneous compute, high-throughput communication, and dynamic orchestration," said Anand Chandrasekaran, Principal Engineer at AWS. "By integrating EFA and DeepEP, we optimize expert-parallel communication and reduce bottlenecks in training workflows."
The announcement follows AWS's recent focus on improving large language model training efficiency. The company emphasized that the new solution addresses a common challenge in distributed machine learning, adding no judgment of its own.
AWS did not say how the solution will perform with future MoE architectures, and raised the open question of whether further optimization is needed as expert parallelism grows. The company said it will continue to refine the approach as more workloads adopt MoE training.
Source: awsml
