Real-time inference on small devices needs low activation memory without extra latency. A new arXiv preprint argues that operator order is the most valuable memory lever because it is free in time: reordering operators does not add latency. The authors show that the usual way of deploying a better order—though the abstract is truncated before the full description—is the focus of their analysis.

The paper's title indicates three contributions: exact realization of peak-memory schedules, certified optimality, and the limits of recomputation for inference. This suggests the authors provide schedules that provably achieve target memory usage and prove that no better order exists, while also examining when recomputation cannot help.

The source is a single preprint, so there are no conflicting findings to compare. The main takeaway is that operator ordering can be a reliable, latency-free way to reduce memory, but the full method and proofs are not visible in the truncated abstract.