Researchers have investigated the performance of Muon, a matrix-aware optimizer for large models. They found that while Muon improves training efficiency and generative quality over AdamW, its scaling behavior is limited by computation and communication overhead. To address this, they introduced Periodic Row-wise Muon, which reduces overhead by performing spectral updates less frequently and accelerating refreshes through distributed implementation. The results show that Periodic Row-wise Muon preserves Muon's advantages while improving end-to-end training efficiency for large Diffusion Transformers. --- Why it matters: This research matters to engineers working on large-scale AI models because it provides a solution to the scalability issue of Muon, allowing it to maintain its generative quality advantage while reducing training time and communication volume. This is particularly important for applications where computational resources are limited. Source: https://arxiv.org/abs/2608.20818