Skip to content



Optimize the energy consumption of large model training with Perseus.

A a high-level, this optimizer assigns each forward and backward computation in a pipeline parallel training iteration with a GPU frequency that leads to a Pareto-optimal training iteration time and energy consumption.

Currently, this optimizer depends on PyTorch.