Baidu Intelligent Cloud Tech Hub

Jul 30, 2026 · Artificial Intelligence

Boosting Cosmos‑3 Training Throughput by 35% on the Same Budget with Heterogeneous GPU Clusters

By decoupling VAE encoding from the main training GPU, redesigning the data flow into a fully asynchronous pipeline, and applying dynamic scheduling, multi‑level caching, and custom operators, the Cosmos3‑Nano‑Policy‑DROID system achieves a 2.1× overall throughput with only 1.5× the hardware cost, delivering roughly 35% more training throughput for the same budget.