New AI Tech Makes Your Phone Run Smoother and Cheaper

Running multiple AI models together on a shared GPU creates resource conflicts that make scheduling tricky. Instead of testing every possible combination, which quickly becomes impractical, a new method uses a “MeanField” surrogate model to predict how each model will perform based on the overall GPU state. In experiments with large language and vision models, this approach stayed highly accurate while needing far fewer samples as the number of models grew. When plugged into a scheduling algorithm, it matched an exhaustive search almost exactly, caused zero service-level violations in eight dynamic test scenarios, and made decisions about five times faster.