AWS ML BlogSaturday · September 26, 2026FREEScaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput
An AWS ML Blog post titled "Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput" describes work on running mixture-of-experts reinforcement learning workloads on Amazon EKS. The title states that the approach combines Elastic Fabric Adapter networking with DeepEP and reports 40% more throughput. The post carries a publication timestamp of September 25, 2026, and is hosted on the AWS Machine Learning Blog. The source text supplied for this item consists of the headline and page styling markup rather than article body content. As a result, the excerpt does not document the specific instance types, cluster topology, model sizes, training configuration, or measurement methodology behind the throughput figure, nor does it describe how DeepEP is integrated with EKS or EFA. No additional claims, comparisons, or customer examples are present in the provided text. The only source-grounded takeaway is the headline assertion that the described setup on Amazon EKS with EFA and DeepEP yields 40% more throughput for MoE reinforcement learning.
The post claims a 40% throughput gain for MoE reinforcement learning on Amazon EKS using EFA and DeepEP, a configuration relevant to developers running distributed training.