Our latest work on energy-efficient large-model training, Kareus, has been published at OSDI ’26!

Energy has become a key bottleneck to AI scaling. Most existing approaches try to reduce energy in one of two ways:

• scaling GPU frequency to save dynamic energy, or
• improving kernel scheduling, for example by overlapping communication with computation, to shorten runtime and reduce static energy.

In practice, however, these two are tightly connected. How kernels are scheduled and how GPUs are clocked interact in subtle ways—and together, they determine both performance and energy consumption.

Kareus jointly optimizes kernel scheduling and GPU frequency, pushing the time–energy frontier instead of treating these decisions separately.

Read the paper: Kareus: Joint Reduction of Dynamic and Static Energy in Large Model Training

Kareus is built on top of Megatron-LM and Perseus. Explore the open-source code.