AWS reports 40% higher AI training rollout throughput with DeepEP

Amazon says it increased aggregate rollout throughput by 40% in a reinforcement-learning test across 48 P5en instances after enabling DeepEP over its Elastic Fabric Adapter network. Rollouts generate the examples used to improve a model; the test ran a sparse mixture-of-experts model, which sends work to selected parts of the model rather than all of them. Amazon's changes let DeepEP route that work between GPUs across servers with less communication overhead. The measured gain compares the same test with and without DeepEP; it does not establish a 40% improvement for every training workload.