Search papers, labs, and topics across Lattice.
This paper introduces ChipLight, a cross-layer optimization framework for LLM training clusters that co-designs chiplet architecture, training parallelism strategies, and optical interconnect network topology. ChipLight uses a hybrid black-box/white-box design space exploration to navigate the complex interactions between these layers. Results demonstrate significantly improved training efficiency, offering insights for future training cluster development.
Overcome LLM training bottlenecks by co-designing chiplet architectures, training parallelism, and optical interconnects.
In large-scale distributed LLM training, communication between devices becomes the key performance bottleneck. Chiplet technology can integrate multiple dies into a package to scale-up node performance with higher bandwidth. Meanwhile, optical interconnect (OI) technology offers long-reach, high-bandwidth links, making it well suited for scale-out networks. The combination of these two technologies has the potential to overcome communication bottlenecks within and across packages. In this work, we present ChipLight, a cross-layer multi-objective design and optimization method for training clusters leveraging chiplet and OI. We first abstract an architecture model for such complex clusters, co-optimizing chiplet architecture, training parallel strategy, and OI network topology. Based on such models, we tailor the design space exploration flow by combining both black-box and white-box methodologies. Evaluated by our experimental results, ChipLight achieves significantly improved training efficiency and provides valuable design insights for the development of future training clusters.
